1 /*
   2  * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "ci/ciUtilities.hpp"
  27 #include "compiler/compileLog.hpp"
  28 #include "ci/ciValueKlass.hpp"
  29 #include "gc/shared/barrierSet.hpp"
  30 #include "gc/shared/c2/barrierSetC2.hpp"
  31 #include "interpreter/interpreter.hpp"
  32 #include "memory/resourceArea.hpp"
  33 #include "opto/addnode.hpp"
  34 #include "opto/castnode.hpp"
  35 #include "opto/convertnode.hpp"
  36 #include "opto/graphKit.hpp"
  37 #include "opto/idealKit.hpp"
  38 #include "opto/intrinsicnode.hpp"
  39 #include "opto/locknode.hpp"
  40 #include "opto/machnode.hpp"
  41 #include "opto/narrowptrnode.hpp"
  42 #include "opto/opaquenode.hpp"
  43 #include "opto/parse.hpp"
  44 #include "opto/rootnode.hpp"
  45 #include "opto/runtime.hpp"
  46 #include "opto/valuetypenode.hpp"
  47 #include "runtime/deoptimization.hpp"
  48 #include "runtime/sharedRuntime.hpp"
  49 
  50 //----------------------------GraphKit-----------------------------------------
  51 // Main utility constructor.
  52 GraphKit::GraphKit(JVMState* jvms, PhaseGVN* gvn)
  53   : Phase(Phase::Parser),
  54     _env(C->env()),
  55     _gvn((gvn != NULL) ? *gvn : *C->initial_gvn()),
  56     _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
  57 {
  58   assert(gvn == NULL || !gvn->is_IterGVN() || gvn->is_IterGVN()->delay_transform(), "delay transform should be enabled");
  59   _exceptions = jvms->map()->next_exception();
  60   if (_exceptions != NULL)  jvms->map()->set_next_exception(NULL);
  61   set_jvms(jvms);
  62 #ifdef ASSERT
  63   if (_gvn.is_IterGVN() != NULL) {
  64     assert(_gvn.is_IterGVN()->delay_transform(), "Transformation must be delayed if IterGVN is used");
  65     // Save the initial size of _for_igvn worklist for verification (see ~GraphKit)
  66     _worklist_size = _gvn.C->for_igvn()->size();
  67   }
  68 #endif
  69 }
  70 
  71 // Private constructor for parser.
  72 GraphKit::GraphKit()
  73   : Phase(Phase::Parser),
  74     _env(C->env()),
  75     _gvn(*C->initial_gvn()),
  76     _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
  77 {
  78   _exceptions = NULL;
  79   set_map(NULL);
  80   debug_only(_sp = -99);
  81   debug_only(set_bci(-99));
  82 }
  83 
  84 
  85 
  86 //---------------------------clean_stack---------------------------------------
  87 // Clear away rubbish from the stack area of the JVM state.
  88 // This destroys any arguments that may be waiting on the stack.
  89 void GraphKit::clean_stack(int from_sp) {
  90   SafePointNode* map      = this->map();
  91   JVMState*      jvms     = this->jvms();
  92   int            stk_size = jvms->stk_size();
  93   int            stkoff   = jvms->stkoff();
  94   Node*          top      = this->top();
  95   for (int i = from_sp; i < stk_size; i++) {
  96     if (map->in(stkoff + i) != top) {
  97       map->set_req(stkoff + i, top);
  98     }
  99   }
 100 }
 101 
 102 
 103 //--------------------------------sync_jvms-----------------------------------
 104 // Make sure our current jvms agrees with our parse state.
 105 JVMState* GraphKit::sync_jvms() const {
 106   JVMState* jvms = this->jvms();
 107   jvms->set_bci(bci());       // Record the new bci in the JVMState
 108   jvms->set_sp(sp());         // Record the new sp in the JVMState
 109   assert(jvms_in_sync(), "jvms is now in sync");
 110   return jvms;
 111 }
 112 
 113 //--------------------------------sync_jvms_for_reexecute---------------------
 114 // Make sure our current jvms agrees with our parse state.  This version
 115 // uses the reexecute_sp for reexecuting bytecodes.
 116 JVMState* GraphKit::sync_jvms_for_reexecute() {
 117   JVMState* jvms = this->jvms();
 118   jvms->set_bci(bci());          // Record the new bci in the JVMState
 119   jvms->set_sp(reexecute_sp());  // Record the new sp in the JVMState
 120   return jvms;
 121 }
 122 
 123 #ifdef ASSERT
 124 bool GraphKit::jvms_in_sync() const {
 125   Parse* parse = is_Parse();
 126   if (parse == NULL) {
 127     if (bci() !=      jvms()->bci())          return false;
 128     if (sp()  != (int)jvms()->sp())           return false;
 129     return true;
 130   }
 131   if (jvms()->method() != parse->method())    return false;
 132   if (jvms()->bci()    != parse->bci())       return false;
 133   int jvms_sp = jvms()->sp();
 134   if (jvms_sp          != parse->sp())        return false;
 135   int jvms_depth = jvms()->depth();
 136   if (jvms_depth       != parse->depth())     return false;
 137   return true;
 138 }
 139 
 140 // Local helper checks for special internal merge points
 141 // used to accumulate and merge exception states.
 142 // They are marked by the region's in(0) edge being the map itself.
 143 // Such merge points must never "escape" into the parser at large,
 144 // until they have been handed to gvn.transform.
 145 static bool is_hidden_merge(Node* reg) {
 146   if (reg == NULL)  return false;
 147   if (reg->is_Phi()) {
 148     reg = reg->in(0);
 149     if (reg == NULL)  return false;
 150   }
 151   return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
 152 }
 153 
 154 void GraphKit::verify_map() const {
 155   if (map() == NULL)  return;  // null map is OK
 156   assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
 157   assert(!map()->has_exceptions(),    "call add_exception_states_from 1st");
 158   assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
 159 }
 160 
 161 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
 162   assert(ex_map->next_exception() == NULL, "not already part of a chain");
 163   assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
 164 }
 165 #endif
 166 
 167 //---------------------------stop_and_kill_map---------------------------------
 168 // Set _map to NULL, signalling a stop to further bytecode execution.
 169 // First smash the current map's control to a constant, to mark it dead.
 170 void GraphKit::stop_and_kill_map() {
 171   SafePointNode* dead_map = stop();
 172   if (dead_map != NULL) {
 173     dead_map->disconnect_inputs(NULL, C); // Mark the map as killed.
 174     assert(dead_map->is_killed(), "must be so marked");
 175   }
 176 }
 177 
 178 
 179 //--------------------------------stopped--------------------------------------
 180 // Tell if _map is NULL, or control is top.
 181 bool GraphKit::stopped() {
 182   if (map() == NULL)           return true;
 183   else if (control() == top()) return true;
 184   else                         return false;
 185 }
 186 
 187 
 188 //-----------------------------has_ex_handler----------------------------------
 189 // Tell if this method or any caller method has exception handlers.
 190 bool GraphKit::has_ex_handler() {
 191   for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
 192     if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
 193       return true;
 194     }
 195   }
 196   return false;
 197 }
 198 
 199 //------------------------------save_ex_oop------------------------------------
 200 // Save an exception without blowing stack contents or other JVM state.
 201 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
 202   assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
 203   ex_map->add_req(ex_oop);
 204   debug_only(verify_exception_state(ex_map));
 205 }
 206 
 207 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
 208   assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
 209   Node* ex_oop = ex_map->in(ex_map->req()-1);
 210   if (clear_it)  ex_map->del_req(ex_map->req()-1);
 211   return ex_oop;
 212 }
 213 
 214 //-----------------------------saved_ex_oop------------------------------------
 215 // Recover a saved exception from its map.
 216 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
 217   return common_saved_ex_oop(ex_map, false);
 218 }
 219 
 220 //--------------------------clear_saved_ex_oop---------------------------------
 221 // Erase a previously saved exception from its map.
 222 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
 223   return common_saved_ex_oop(ex_map, true);
 224 }
 225 
 226 #ifdef ASSERT
 227 //---------------------------has_saved_ex_oop----------------------------------
 228 // Erase a previously saved exception from its map.
 229 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
 230   return ex_map->req() == ex_map->jvms()->endoff()+1;
 231 }
 232 #endif
 233 
 234 //-------------------------make_exception_state--------------------------------
 235 // Turn the current JVM state into an exception state, appending the ex_oop.
 236 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
 237   sync_jvms();
 238   SafePointNode* ex_map = stop();  // do not manipulate this map any more
 239   set_saved_ex_oop(ex_map, ex_oop);
 240   return ex_map;
 241 }
 242 
 243 
 244 //--------------------------add_exception_state--------------------------------
 245 // Add an exception to my list of exceptions.
 246 void GraphKit::add_exception_state(SafePointNode* ex_map) {
 247   if (ex_map == NULL || ex_map->control() == top()) {
 248     return;
 249   }
 250 #ifdef ASSERT
 251   verify_exception_state(ex_map);
 252   if (has_exceptions()) {
 253     assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
 254   }
 255 #endif
 256 
 257   // If there is already an exception of exactly this type, merge with it.
 258   // In particular, null-checks and other low-level exceptions common up here.
 259   Node*       ex_oop  = saved_ex_oop(ex_map);
 260   const Type* ex_type = _gvn.type(ex_oop);
 261   if (ex_oop == top()) {
 262     // No action needed.
 263     return;
 264   }
 265   assert(ex_type->isa_instptr(), "exception must be an instance");
 266   for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
 267     const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
 268     // We check sp also because call bytecodes can generate exceptions
 269     // both before and after arguments are popped!
 270     if (ex_type2 == ex_type
 271         && e2->_jvms->sp() == ex_map->_jvms->sp()) {
 272       combine_exception_states(ex_map, e2);
 273       return;
 274     }
 275   }
 276 
 277   // No pre-existing exception of the same type.  Chain it on the list.
 278   push_exception_state(ex_map);
 279 }
 280 
 281 //-----------------------add_exception_states_from-----------------------------
 282 void GraphKit::add_exception_states_from(JVMState* jvms) {
 283   SafePointNode* ex_map = jvms->map()->next_exception();
 284   if (ex_map != NULL) {
 285     jvms->map()->set_next_exception(NULL);
 286     for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
 287       next_map = ex_map->next_exception();
 288       ex_map->set_next_exception(NULL);
 289       add_exception_state(ex_map);
 290     }
 291   }
 292 }
 293 
 294 //-----------------------transfer_exceptions_into_jvms-------------------------
 295 JVMState* GraphKit::transfer_exceptions_into_jvms() {
 296   if (map() == NULL) {
 297     // We need a JVMS to carry the exceptions, but the map has gone away.
 298     // Create a scratch JVMS, cloned from any of the exception states...
 299     if (has_exceptions()) {
 300       _map = _exceptions;
 301       _map = clone_map();
 302       _map->set_next_exception(NULL);
 303       clear_saved_ex_oop(_map);
 304       debug_only(verify_map());
 305     } else {
 306       // ...or created from scratch
 307       JVMState* jvms = new (C) JVMState(_method, NULL);
 308       jvms->set_bci(_bci);
 309       jvms->set_sp(_sp);
 310       jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms));
 311       set_jvms(jvms);
 312       for (uint i = 0; i < map()->req(); i++)  map()->init_req(i, top());
 313       set_all_memory(top());
 314       while (map()->req() < jvms->endoff())  map()->add_req(top());
 315     }
 316     // (This is a kludge, in case you didn't notice.)
 317     set_control(top());
 318   }
 319   JVMState* jvms = sync_jvms();
 320   assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
 321   jvms->map()->set_next_exception(_exceptions);
 322   _exceptions = NULL;   // done with this set of exceptions
 323   return jvms;
 324 }
 325 
 326 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
 327   assert(is_hidden_merge(dstphi), "must be a special merge node");
 328   assert(is_hidden_merge(srcphi), "must be a special merge node");
 329   uint limit = srcphi->req();
 330   for (uint i = PhiNode::Input; i < limit; i++) {
 331     dstphi->add_req(srcphi->in(i));
 332   }
 333 }
 334 static inline void add_one_req(Node* dstphi, Node* src) {
 335   assert(is_hidden_merge(dstphi), "must be a special merge node");
 336   assert(!is_hidden_merge(src), "must not be a special merge node");
 337   dstphi->add_req(src);
 338 }
 339 
 340 //-----------------------combine_exception_states------------------------------
 341 // This helper function combines exception states by building phis on a
 342 // specially marked state-merging region.  These regions and phis are
 343 // untransformed, and can build up gradually.  The region is marked by
 344 // having a control input of its exception map, rather than NULL.  Such
 345 // regions do not appear except in this function, and in use_exception_state.
 346 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
 347   if (failing())  return;  // dying anyway...
 348   JVMState* ex_jvms = ex_map->_jvms;
 349   assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
 350   assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
 351   assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
 352   assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
 353   assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
 354   assert(ex_map->req() == phi_map->req(), "matching maps");
 355   uint tos = ex_jvms->stkoff() + ex_jvms->sp();
 356   Node*         hidden_merge_mark = root();
 357   Node*         region  = phi_map->control();
 358   MergeMemNode* phi_mem = phi_map->merged_memory();
 359   MergeMemNode* ex_mem  = ex_map->merged_memory();
 360   if (region->in(0) != hidden_merge_mark) {
 361     // The control input is not (yet) a specially-marked region in phi_map.
 362     // Make it so, and build some phis.
 363     region = new RegionNode(2);
 364     _gvn.set_type(region, Type::CONTROL);
 365     region->set_req(0, hidden_merge_mark);  // marks an internal ex-state
 366     region->init_req(1, phi_map->control());
 367     phi_map->set_control(region);
 368     Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
 369     record_for_igvn(io_phi);
 370     _gvn.set_type(io_phi, Type::ABIO);
 371     phi_map->set_i_o(io_phi);
 372     for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
 373       Node* m = mms.memory();
 374       Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
 375       record_for_igvn(m_phi);
 376       _gvn.set_type(m_phi, Type::MEMORY);
 377       mms.set_memory(m_phi);
 378     }
 379   }
 380 
 381   // Either or both of phi_map and ex_map might already be converted into phis.
 382   Node* ex_control = ex_map->control();
 383   // if there is special marking on ex_map also, we add multiple edges from src
 384   bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
 385   // how wide was the destination phi_map, originally?
 386   uint orig_width = region->req();
 387 
 388   if (add_multiple) {
 389     add_n_reqs(region, ex_control);
 390     add_n_reqs(phi_map->i_o(), ex_map->i_o());
 391   } else {
 392     // ex_map has no merges, so we just add single edges everywhere
 393     add_one_req(region, ex_control);
 394     add_one_req(phi_map->i_o(), ex_map->i_o());
 395   }
 396   for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
 397     if (mms.is_empty()) {
 398       // get a copy of the base memory, and patch some inputs into it
 399       const TypePtr* adr_type = mms.adr_type(C);
 400       Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
 401       assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
 402       mms.set_memory(phi);
 403       // Prepare to append interesting stuff onto the newly sliced phi:
 404       while (phi->req() > orig_width)  phi->del_req(phi->req()-1);
 405     }
 406     // Append stuff from ex_map:
 407     if (add_multiple) {
 408       add_n_reqs(mms.memory(), mms.memory2());
 409     } else {
 410       add_one_req(mms.memory(), mms.memory2());
 411     }
 412   }
 413   uint limit = ex_map->req();
 414   for (uint i = TypeFunc::Parms; i < limit; i++) {
 415     // Skip everything in the JVMS after tos.  (The ex_oop follows.)
 416     if (i == tos)  i = ex_jvms->monoff();
 417     Node* src = ex_map->in(i);
 418     Node* dst = phi_map->in(i);
 419     if (src != dst) {
 420       PhiNode* phi;
 421       if (dst->in(0) != region) {
 422         dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
 423         record_for_igvn(phi);
 424         _gvn.set_type(phi, phi->type());
 425         phi_map->set_req(i, dst);
 426         // Prepare to append interesting stuff onto the new phi:
 427         while (dst->req() > orig_width)  dst->del_req(dst->req()-1);
 428       } else {
 429         assert(dst->is_Phi(), "nobody else uses a hidden region");
 430         phi = dst->as_Phi();
 431       }
 432       if (add_multiple && src->in(0) == ex_control) {
 433         // Both are phis.
 434         add_n_reqs(dst, src);
 435       } else {
 436         while (dst->req() < region->req())  add_one_req(dst, src);
 437       }
 438       const Type* srctype = _gvn.type(src);
 439       if (phi->type() != srctype) {
 440         const Type* dsttype = phi->type()->meet_speculative(srctype);
 441         if (phi->type() != dsttype) {
 442           phi->set_type(dsttype);
 443           _gvn.set_type(phi, dsttype);
 444         }
 445       }
 446     }
 447   }
 448   phi_map->merge_replaced_nodes_with(ex_map);
 449 }
 450 
 451 //--------------------------use_exception_state--------------------------------
 452 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
 453   if (failing()) { stop(); return top(); }
 454   Node* region = phi_map->control();
 455   Node* hidden_merge_mark = root();
 456   assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
 457   Node* ex_oop = clear_saved_ex_oop(phi_map);
 458   if (region->in(0) == hidden_merge_mark) {
 459     // Special marking for internal ex-states.  Process the phis now.
 460     region->set_req(0, region);  // now it's an ordinary region
 461     set_jvms(phi_map->jvms());   // ...so now we can use it as a map
 462     // Note: Setting the jvms also sets the bci and sp.
 463     set_control(_gvn.transform(region));
 464     uint tos = jvms()->stkoff() + sp();
 465     for (uint i = 1; i < tos; i++) {
 466       Node* x = phi_map->in(i);
 467       if (x->in(0) == region) {
 468         assert(x->is_Phi(), "expected a special phi");
 469         phi_map->set_req(i, _gvn.transform(x));
 470       }
 471     }
 472     for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
 473       Node* x = mms.memory();
 474       if (x->in(0) == region) {
 475         assert(x->is_Phi(), "nobody else uses a hidden region");
 476         mms.set_memory(_gvn.transform(x));
 477       }
 478     }
 479     if (ex_oop->in(0) == region) {
 480       assert(ex_oop->is_Phi(), "expected a special phi");
 481       ex_oop = _gvn.transform(ex_oop);
 482     }
 483   } else {
 484     set_jvms(phi_map->jvms());
 485   }
 486 
 487   assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
 488   assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
 489   return ex_oop;
 490 }
 491 
 492 //---------------------------------java_bc-------------------------------------
 493 Bytecodes::Code GraphKit::java_bc() const {
 494   ciMethod* method = this->method();
 495   int       bci    = this->bci();
 496   if (method != NULL && bci != InvocationEntryBci)
 497     return method->java_code_at_bci(bci);
 498   else
 499     return Bytecodes::_illegal;
 500 }
 501 
 502 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
 503                                                           bool must_throw) {
 504     // if the exception capability is set, then we will generate code
 505     // to check the JavaThread.should_post_on_exceptions flag to see
 506     // if we actually need to report exception events (for this
 507     // thread).  If we don't need to report exception events, we will
 508     // take the normal fast path provided by add_exception_events.  If
 509     // exception event reporting is enabled for this thread, we will
 510     // take the uncommon_trap in the BuildCutout below.
 511 
 512     // first must access the should_post_on_exceptions_flag in this thread's JavaThread
 513     Node* jthread = _gvn.transform(new ThreadLocalNode());
 514     Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
 515     Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
 516 
 517     // Test the should_post_on_exceptions_flag vs. 0
 518     Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) );
 519     Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) );
 520 
 521     // Branch to slow_path if should_post_on_exceptions_flag was true
 522     { BuildCutout unless(this, tst, PROB_MAX);
 523       // Do not try anything fancy if we're notifying the VM on every throw.
 524       // Cf. case Bytecodes::_athrow in parse2.cpp.
 525       uncommon_trap(reason, Deoptimization::Action_none,
 526                     (ciKlass*)NULL, (char*)NULL, must_throw);
 527     }
 528 
 529 }
 530 
 531 //------------------------------builtin_throw----------------------------------
 532 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
 533   bool must_throw = true;
 534 
 535   if (env()->jvmti_can_post_on_exceptions()) {
 536     // check if we must post exception events, take uncommon trap if so
 537     uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
 538     // here if should_post_on_exceptions is false
 539     // continue on with the normal codegen
 540   }
 541 
 542   // If this particular condition has not yet happened at this
 543   // bytecode, then use the uncommon trap mechanism, and allow for
 544   // a future recompilation if several traps occur here.
 545   // If the throw is hot, try to use a more complicated inline mechanism
 546   // which keeps execution inside the compiled code.
 547   bool treat_throw_as_hot = false;
 548   ciMethodData* md = method()->method_data();
 549 
 550   if (ProfileTraps) {
 551     if (too_many_traps(reason)) {
 552       treat_throw_as_hot = true;
 553     }
 554     // (If there is no MDO at all, assume it is early in
 555     // execution, and that any deopts are part of the
 556     // startup transient, and don't need to be remembered.)
 557 
 558     // Also, if there is a local exception handler, treat all throws
 559     // as hot if there has been at least one in this method.
 560     if (C->trap_count(reason) != 0
 561         && method()->method_data()->trap_count(reason) != 0
 562         && has_ex_handler()) {
 563         treat_throw_as_hot = true;
 564     }
 565   }
 566 
 567   // If this throw happens frequently, an uncommon trap might cause
 568   // a performance pothole.  If there is a local exception handler,
 569   // and if this particular bytecode appears to be deoptimizing often,
 570   // let us handle the throw inline, with a preconstructed instance.
 571   // Note:   If the deopt count has blown up, the uncommon trap
 572   // runtime is going to flush this nmethod, not matter what.
 573   if (treat_throw_as_hot
 574       && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
 575     // If the throw is local, we use a pre-existing instance and
 576     // punt on the backtrace.  This would lead to a missing backtrace
 577     // (a repeat of 4292742) if the backtrace object is ever asked
 578     // for its backtrace.
 579     // Fixing this remaining case of 4292742 requires some flavor of
 580     // escape analysis.  Leave that for the future.
 581     ciInstance* ex_obj = NULL;
 582     switch (reason) {
 583     case Deoptimization::Reason_null_check:
 584       ex_obj = env()->NullPointerException_instance();
 585       break;
 586     case Deoptimization::Reason_div0_check:
 587       ex_obj = env()->ArithmeticException_instance();
 588       break;
 589     case Deoptimization::Reason_range_check:
 590       ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
 591       break;
 592     case Deoptimization::Reason_class_check:
 593       if (java_bc() == Bytecodes::_aastore) {
 594         ex_obj = env()->ArrayStoreException_instance();
 595       } else {
 596         ex_obj = env()->ClassCastException_instance();
 597       }
 598       break;
 599     default:
 600       break;
 601     }
 602     if (failing()) { stop(); return; }  // exception allocation might fail
 603     if (ex_obj != NULL) {
 604       // Cheat with a preallocated exception object.
 605       if (C->log() != NULL)
 606         C->log()->elem("hot_throw preallocated='1' reason='%s'",
 607                        Deoptimization::trap_reason_name(reason));
 608       const TypeInstPtr* ex_con  = TypeInstPtr::make(ex_obj);
 609       Node*              ex_node = _gvn.transform(ConNode::make(ex_con));
 610 
 611       // Clear the detail message of the preallocated exception object.
 612       // Weblogic sometimes mutates the detail message of exceptions
 613       // using reflection.
 614       int offset = java_lang_Throwable::get_detailMessage_offset();
 615       const TypePtr* adr_typ = ex_con->add_offset(offset);
 616 
 617       Node *adr = basic_plus_adr(ex_node, ex_node, offset);
 618       const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
 619       Node *store = access_store_at(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, IN_HEAP);
 620 
 621       add_exception_state(make_exception_state(ex_node));
 622       return;
 623     }
 624   }
 625 
 626   // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
 627   // It won't be much cheaper than bailing to the interp., since we'll
 628   // have to pass up all the debug-info, and the runtime will have to
 629   // create the stack trace.
 630 
 631   // Usual case:  Bail to interpreter.
 632   // Reserve the right to recompile if we haven't seen anything yet.
 633 
 634   ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL;
 635   Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
 636   if (treat_throw_as_hot
 637       && (method()->method_data()->trap_recompiled_at(bci(), m)
 638           || C->too_many_traps(reason))) {
 639     // We cannot afford to take more traps here.  Suffer in the interpreter.
 640     if (C->log() != NULL)
 641       C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
 642                      Deoptimization::trap_reason_name(reason),
 643                      C->trap_count(reason));
 644     action = Deoptimization::Action_none;
 645   }
 646 
 647   // "must_throw" prunes the JVM state to include only the stack, if there
 648   // are no local exception handlers.  This should cut down on register
 649   // allocation time and code size, by drastically reducing the number
 650   // of in-edges on the call to the uncommon trap.
 651 
 652   uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
 653 }
 654 
 655 
 656 //----------------------------PreserveJVMState---------------------------------
 657 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
 658   debug_only(kit->verify_map());
 659   _kit    = kit;
 660   _map    = kit->map();   // preserve the map
 661   _sp     = kit->sp();
 662   kit->set_map(clone_map ? kit->clone_map() : NULL);
 663 #ifdef ASSERT
 664   _bci    = kit->bci();
 665   Parse* parser = kit->is_Parse();
 666   int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
 667   _block  = block;
 668 #endif
 669 }
 670 PreserveJVMState::~PreserveJVMState() {
 671   GraphKit* kit = _kit;
 672 #ifdef ASSERT
 673   assert(kit->bci() == _bci, "bci must not shift");
 674   Parse* parser = kit->is_Parse();
 675   int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
 676   assert(block == _block,    "block must not shift");
 677 #endif
 678   kit->set_map(_map);
 679   kit->set_sp(_sp);
 680 }
 681 
 682 
 683 //-----------------------------BuildCutout-------------------------------------
 684 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
 685   : PreserveJVMState(kit)
 686 {
 687   assert(p->is_Con() || p->is_Bool(), "test must be a bool");
 688   SafePointNode* outer_map = _map;   // preserved map is caller's
 689   SafePointNode* inner_map = kit->map();
 690   IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
 691   outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) ));
 692   inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) ));
 693 }
 694 BuildCutout::~BuildCutout() {
 695   GraphKit* kit = _kit;
 696   assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
 697 }
 698 
 699 //---------------------------PreserveReexecuteState----------------------------
 700 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
 701   assert(!kit->stopped(), "must call stopped() before");
 702   _kit    =    kit;
 703   _sp     =    kit->sp();
 704   _reexecute = kit->jvms()->_reexecute;
 705 }
 706 PreserveReexecuteState::~PreserveReexecuteState() {
 707   if (_kit->stopped()) return;
 708   _kit->jvms()->_reexecute = _reexecute;
 709   _kit->set_sp(_sp);
 710 }
 711 
 712 //------------------------------clone_map--------------------------------------
 713 // Implementation of PreserveJVMState
 714 //
 715 // Only clone_map(...) here. If this function is only used in the
 716 // PreserveJVMState class we may want to get rid of this extra
 717 // function eventually and do it all there.
 718 
 719 SafePointNode* GraphKit::clone_map() {
 720   if (map() == NULL)  return NULL;
 721 
 722   // Clone the memory edge first
 723   Node* mem = MergeMemNode::make(map()->memory());
 724   gvn().set_type_bottom(mem);
 725 
 726   SafePointNode *clonemap = (SafePointNode*)map()->clone();
 727   JVMState* jvms = this->jvms();
 728   JVMState* clonejvms = jvms->clone_shallow(C);
 729   clonemap->set_memory(mem);
 730   clonemap->set_jvms(clonejvms);
 731   clonejvms->set_map(clonemap);
 732   record_for_igvn(clonemap);
 733   gvn().set_type_bottom(clonemap);
 734   return clonemap;
 735 }
 736 
 737 
 738 //-----------------------------set_map_clone-----------------------------------
 739 void GraphKit::set_map_clone(SafePointNode* m) {
 740   _map = m;
 741   _map = clone_map();
 742   _map->set_next_exception(NULL);
 743   debug_only(verify_map());
 744 }
 745 
 746 
 747 //----------------------------kill_dead_locals---------------------------------
 748 // Detect any locals which are known to be dead, and force them to top.
 749 void GraphKit::kill_dead_locals() {
 750   // Consult the liveness information for the locals.  If any
 751   // of them are unused, then they can be replaced by top().  This
 752   // should help register allocation time and cut down on the size
 753   // of the deoptimization information.
 754 
 755   // This call is made from many of the bytecode handling
 756   // subroutines called from the Big Switch in do_one_bytecode.
 757   // Every bytecode which might include a slow path is responsible
 758   // for killing its dead locals.  The more consistent we
 759   // are about killing deads, the fewer useless phis will be
 760   // constructed for them at various merge points.
 761 
 762   // bci can be -1 (InvocationEntryBci).  We return the entry
 763   // liveness for the method.
 764 
 765   if (method() == NULL || method()->code_size() == 0) {
 766     // We are building a graph for a call to a native method.
 767     // All locals are live.
 768     return;
 769   }
 770 
 771   ResourceMark rm;
 772 
 773   // Consult the liveness information for the locals.  If any
 774   // of them are unused, then they can be replaced by top().  This
 775   // should help register allocation time and cut down on the size
 776   // of the deoptimization information.
 777   MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
 778 
 779   int len = (int)live_locals.size();
 780   assert(len <= jvms()->loc_size(), "too many live locals");
 781   for (int local = 0; local < len; local++) {
 782     if (!live_locals.at(local)) {
 783       set_local(local, top());
 784     }
 785   }
 786 }
 787 
 788 #ifdef ASSERT
 789 //-------------------------dead_locals_are_killed------------------------------
 790 // Return true if all dead locals are set to top in the map.
 791 // Used to assert "clean" debug info at various points.
 792 bool GraphKit::dead_locals_are_killed() {
 793   if (method() == NULL || method()->code_size() == 0) {
 794     // No locals need to be dead, so all is as it should be.
 795     return true;
 796   }
 797 
 798   // Make sure somebody called kill_dead_locals upstream.
 799   ResourceMark rm;
 800   for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
 801     if (jvms->loc_size() == 0)  continue;  // no locals to consult
 802     SafePointNode* map = jvms->map();
 803     ciMethod* method = jvms->method();
 804     int       bci    = jvms->bci();
 805     if (jvms == this->jvms()) {
 806       bci = this->bci();  // it might not yet be synched
 807     }
 808     MethodLivenessResult live_locals = method->liveness_at_bci(bci);
 809     int len = (int)live_locals.size();
 810     if (!live_locals.is_valid() || len == 0)
 811       // This method is trivial, or is poisoned by a breakpoint.
 812       return true;
 813     assert(len == jvms->loc_size(), "live map consistent with locals map");
 814     for (int local = 0; local < len; local++) {
 815       if (!live_locals.at(local) && map->local(jvms, local) != top()) {
 816         if (PrintMiscellaneous && (Verbose || WizardMode)) {
 817           tty->print_cr("Zombie local %d: ", local);
 818           jvms->dump();
 819         }
 820         return false;
 821       }
 822     }
 823   }
 824   return true;
 825 }
 826 
 827 #endif //ASSERT
 828 
 829 // Helper function for enforcing certain bytecodes to reexecute if
 830 // deoptimization happens
 831 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
 832   ciMethod* cur_method = jvms->method();
 833   int       cur_bci   = jvms->bci();
 834   if (cur_method != NULL && cur_bci != InvocationEntryBci) {
 835     Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
 836     return Interpreter::bytecode_should_reexecute(code) ||
 837            (is_anewarray && (code == Bytecodes::_multianewarray));
 838     // Reexecute _multianewarray bytecode which was replaced with
 839     // sequence of [a]newarray. See Parse::do_multianewarray().
 840     //
 841     // Note: interpreter should not have it set since this optimization
 842     // is limited by dimensions and guarded by flag so in some cases
 843     // multianewarray() runtime calls will be generated and
 844     // the bytecode should not be reexecutes (stack will not be reset).
 845   } else {
 846     return false;
 847   }
 848 }
 849 
 850 // Helper function for adding JVMState and debug information to node
 851 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
 852   // Add the safepoint edges to the call (or other safepoint).
 853 
 854   // Make sure dead locals are set to top.  This
 855   // should help register allocation time and cut down on the size
 856   // of the deoptimization information.
 857   assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
 858 
 859   // Walk the inline list to fill in the correct set of JVMState's
 860   // Also fill in the associated edges for each JVMState.
 861 
 862   // If the bytecode needs to be reexecuted we need to put
 863   // the arguments back on the stack.
 864   const bool should_reexecute = jvms()->should_reexecute();
 865   JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
 866 
 867   // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
 868   // undefined if the bci is different.  This is normal for Parse but it
 869   // should not happen for LibraryCallKit because only one bci is processed.
 870   assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
 871          "in LibraryCallKit the reexecute bit should not change");
 872 
 873   // If we are guaranteed to throw, we can prune everything but the
 874   // input to the current bytecode.
 875   bool can_prune_locals = false;
 876   uint stack_slots_not_pruned = 0;
 877   int inputs = 0, depth = 0;
 878   if (must_throw) {
 879     assert(method() == youngest_jvms->method(), "sanity");
 880     if (compute_stack_effects(inputs, depth)) {
 881       can_prune_locals = true;
 882       stack_slots_not_pruned = inputs;
 883     }
 884   }
 885 
 886   if (env()->should_retain_local_variables()) {
 887     // At any safepoint, this method can get breakpointed, which would
 888     // then require an immediate deoptimization.
 889     can_prune_locals = false;  // do not prune locals
 890     stack_slots_not_pruned = 0;
 891   }
 892 
 893   // do not scribble on the input jvms
 894   JVMState* out_jvms = youngest_jvms->clone_deep(C);
 895   call->set_jvms(out_jvms); // Start jvms list for call node
 896 
 897   // For a known set of bytecodes, the interpreter should reexecute them if
 898   // deoptimization happens. We set the reexecute state for them here
 899   if (out_jvms->is_reexecute_undefined() && //don't change if already specified
 900       should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
 901     out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
 902   }
 903 
 904   // Presize the call:
 905   DEBUG_ONLY(uint non_debug_edges = call->req());
 906   call->add_req_batch(top(), youngest_jvms->debug_depth());
 907   assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
 908 
 909   // Set up edges so that the call looks like this:
 910   //  Call [state:] ctl io mem fptr retadr
 911   //       [parms:] parm0 ... parmN
 912   //       [root:]  loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
 913   //    [...mid:]   loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
 914   //       [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
 915   // Note that caller debug info precedes callee debug info.
 916 
 917   // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
 918   uint debug_ptr = call->req();
 919 
 920   // Loop over the map input edges associated with jvms, add them
 921   // to the call node, & reset all offsets to match call node array.
 922   for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
 923     uint debug_end   = debug_ptr;
 924     uint debug_start = debug_ptr - in_jvms->debug_size();
 925     debug_ptr = debug_start;  // back up the ptr
 926 
 927     uint p = debug_start;  // walks forward in [debug_start, debug_end)
 928     uint j, k, l;
 929     SafePointNode* in_map = in_jvms->map();
 930     out_jvms->set_map(call);
 931 
 932     if (can_prune_locals) {
 933       assert(in_jvms->method() == out_jvms->method(), "sanity");
 934       // If the current throw can reach an exception handler in this JVMS,
 935       // then we must keep everything live that can reach that handler.
 936       // As a quick and dirty approximation, we look for any handlers at all.
 937       if (in_jvms->method()->has_exception_handlers()) {
 938         can_prune_locals = false;
 939       }
 940     }
 941 
 942     // Add the Locals
 943     k = in_jvms->locoff();
 944     l = in_jvms->loc_size();
 945     out_jvms->set_locoff(p);
 946     if (!can_prune_locals) {
 947       for (j = 0; j < l; j++)
 948         call->set_req(p++, in_map->in(k+j));
 949     } else {
 950       p += l;  // already set to top above by add_req_batch
 951     }
 952 
 953     // Add the Expression Stack
 954     k = in_jvms->stkoff();
 955     l = in_jvms->sp();
 956     out_jvms->set_stkoff(p);
 957     if (!can_prune_locals) {
 958       for (j = 0; j < l; j++)
 959         call->set_req(p++, in_map->in(k+j));
 960     } else if (can_prune_locals && stack_slots_not_pruned != 0) {
 961       // Divide stack into {S0,...,S1}, where S0 is set to top.
 962       uint s1 = stack_slots_not_pruned;
 963       stack_slots_not_pruned = 0;  // for next iteration
 964       if (s1 > l)  s1 = l;
 965       uint s0 = l - s1;
 966       p += s0;  // skip the tops preinstalled by add_req_batch
 967       for (j = s0; j < l; j++)
 968         call->set_req(p++, in_map->in(k+j));
 969     } else {
 970       p += l;  // already set to top above by add_req_batch
 971     }
 972 
 973     // Add the Monitors
 974     k = in_jvms->monoff();
 975     l = in_jvms->mon_size();
 976     out_jvms->set_monoff(p);
 977     for (j = 0; j < l; j++)
 978       call->set_req(p++, in_map->in(k+j));
 979 
 980     // Copy any scalar object fields.
 981     k = in_jvms->scloff();
 982     l = in_jvms->scl_size();
 983     out_jvms->set_scloff(p);
 984     for (j = 0; j < l; j++)
 985       call->set_req(p++, in_map->in(k+j));
 986 
 987     // Finish the new jvms.
 988     out_jvms->set_endoff(p);
 989 
 990     assert(out_jvms->endoff()     == debug_end,             "fill ptr must match");
 991     assert(out_jvms->depth()      == in_jvms->depth(),      "depth must match");
 992     assert(out_jvms->loc_size()   == in_jvms->loc_size(),   "size must match");
 993     assert(out_jvms->mon_size()   == in_jvms->mon_size(),   "size must match");
 994     assert(out_jvms->scl_size()   == in_jvms->scl_size(),   "size must match");
 995     assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
 996 
 997     // Update the two tail pointers in parallel.
 998     out_jvms = out_jvms->caller();
 999     in_jvms  = in_jvms->caller();
1000   }
1001 
1002   assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
1003 
1004   // Test the correctness of JVMState::debug_xxx accessors:
1005   assert(call->jvms()->debug_start() == non_debug_edges, "");
1006   assert(call->jvms()->debug_end()   == call->req(), "");
1007   assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1008 }
1009 
1010 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1011   Bytecodes::Code code = java_bc();
1012   if (code == Bytecodes::_wide) {
1013     code = method()->java_code_at_bci(bci() + 1);
1014   }
1015 
1016   BasicType rtype = T_ILLEGAL;
1017   int       rsize = 0;
1018 
1019   if (code != Bytecodes::_illegal) {
1020     depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1021     rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1022     if (rtype < T_CONFLICT)
1023       rsize = type2size[rtype];
1024   }
1025 
1026   switch (code) {
1027   case Bytecodes::_illegal:
1028     return false;
1029 
1030   case Bytecodes::_ldc:
1031   case Bytecodes::_ldc_w:
1032   case Bytecodes::_ldc2_w:
1033     inputs = 0;
1034     break;
1035 
1036   case Bytecodes::_dup:         inputs = 1;  break;
1037   case Bytecodes::_dup_x1:      inputs = 2;  break;
1038   case Bytecodes::_dup_x2:      inputs = 3;  break;
1039   case Bytecodes::_dup2:        inputs = 2;  break;
1040   case Bytecodes::_dup2_x1:     inputs = 3;  break;
1041   case Bytecodes::_dup2_x2:     inputs = 4;  break;
1042   case Bytecodes::_swap:        inputs = 2;  break;
1043   case Bytecodes::_arraylength: inputs = 1;  break;
1044 
1045   case Bytecodes::_getstatic:
1046   case Bytecodes::_putstatic:
1047   case Bytecodes::_getfield:
1048   case Bytecodes::_putfield:
1049     {
1050       bool ignored_will_link;
1051       ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1052       int      size  = field->type()->size();
1053       bool is_get = (depth >= 0), is_static = (depth & 1);
1054       inputs = (is_static ? 0 : 1);
1055       if (is_get) {
1056         depth = size - inputs;
1057       } else {
1058         inputs += size;        // putxxx pops the value from the stack
1059         depth = - inputs;
1060       }
1061     }
1062     break;
1063 
1064   case Bytecodes::_invokevirtual:
1065   case Bytecodes::_invokespecial:
1066   case Bytecodes::_invokestatic:
1067   case Bytecodes::_invokedynamic:
1068   case Bytecodes::_invokeinterface:
1069     {
1070       bool ignored_will_link;
1071       ciSignature* declared_signature = NULL;
1072       ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1073       assert(declared_signature != NULL, "cannot be null");
1074       inputs   = declared_signature->arg_size_for_bc(code);
1075       int size = declared_signature->return_type()->size();
1076       depth = size - inputs;
1077     }
1078     break;
1079 
1080   case Bytecodes::_multianewarray:
1081     {
1082       ciBytecodeStream iter(method());
1083       iter.reset_to_bci(bci());
1084       iter.next();
1085       inputs = iter.get_dimensions();
1086       assert(rsize == 1, "");
1087       depth = rsize - inputs;
1088     }
1089     break;
1090 
1091   case Bytecodes::_withfield: {
1092     bool ignored_will_link;
1093     ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1094     int      size  = field->type()->size();
1095     inputs = size+1;
1096     depth = rsize - inputs;
1097     break;
1098   }
1099 
1100   case Bytecodes::_ireturn:
1101   case Bytecodes::_lreturn:
1102   case Bytecodes::_freturn:
1103   case Bytecodes::_dreturn:
1104   case Bytecodes::_areturn:
1105     assert(rsize == -depth, "");
1106     inputs = rsize;
1107     break;
1108 
1109   case Bytecodes::_jsr:
1110   case Bytecodes::_jsr_w:
1111     inputs = 0;
1112     depth  = 1;                  // S.B. depth=1, not zero
1113     break;
1114 
1115   default:
1116     // bytecode produces a typed result
1117     inputs = rsize - depth;
1118     assert(inputs >= 0, "");
1119     break;
1120   }
1121 
1122 #ifdef ASSERT
1123   // spot check
1124   int outputs = depth + inputs;
1125   assert(outputs >= 0, "sanity");
1126   switch (code) {
1127   case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1128   case Bytecodes::_athrow:    assert(inputs == 1 && outputs == 0, ""); break;
1129   case Bytecodes::_aload_0:   assert(inputs == 0 && outputs == 1, ""); break;
1130   case Bytecodes::_return:    assert(inputs == 0 && outputs == 0, ""); break;
1131   case Bytecodes::_drem:      assert(inputs == 4 && outputs == 2, ""); break;
1132   default:                    break;
1133   }
1134 #endif //ASSERT
1135 
1136   return true;
1137 }
1138 
1139 
1140 
1141 //------------------------------basic_plus_adr---------------------------------
1142 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1143   // short-circuit a common case
1144   if (offset == intcon(0))  return ptr;
1145   return _gvn.transform( new AddPNode(base, ptr, offset) );
1146 }
1147 
1148 Node* GraphKit::ConvI2L(Node* offset) {
1149   // short-circuit a common case
1150   jint offset_con = find_int_con(offset, Type::OffsetBot);
1151   if (offset_con != Type::OffsetBot) {
1152     return longcon((jlong) offset_con);
1153   }
1154   return _gvn.transform( new ConvI2LNode(offset));
1155 }
1156 
1157 Node* GraphKit::ConvI2UL(Node* offset) {
1158   juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1159   if (offset_con != (juint) Type::OffsetBot) {
1160     return longcon((julong) offset_con);
1161   }
1162   Node* conv = _gvn.transform( new ConvI2LNode(offset));
1163   Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1164   return _gvn.transform( new AndLNode(conv, mask) );
1165 }
1166 
1167 Node* GraphKit::ConvL2I(Node* offset) {
1168   // short-circuit a common case
1169   jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1170   if (offset_con != (jlong)Type::OffsetBot) {
1171     return intcon((int) offset_con);
1172   }
1173   return _gvn.transform( new ConvL2INode(offset));
1174 }
1175 
1176 //-------------------------load_object_klass-----------------------------------
1177 Node* GraphKit::load_object_klass(Node* obj) {
1178   // Special-case a fresh allocation to avoid building nodes:
1179   Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1180   if (akls != NULL)  return akls;
1181   Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1182   return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1183 }
1184 
1185 //-------------------------load_array_length-----------------------------------
1186 Node* GraphKit::load_array_length(Node* array) {
1187   // Special-case a fresh allocation to avoid building nodes:
1188   AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1189   Node *alen;
1190   if (alloc == NULL) {
1191     Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1192     alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1193   } else {
1194     alen = alloc->Ideal_length();
1195     Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1196     if (ccast != alen) {
1197       alen = _gvn.transform(ccast);
1198     }
1199   }
1200   return alen;
1201 }
1202 
1203 //------------------------------do_null_check----------------------------------
1204 // Helper function to do a NULL pointer check.  Returned value is
1205 // the incoming address with NULL casted away.  You are allowed to use the
1206 // not-null value only if you are control dependent on the test.
1207 #ifndef PRODUCT
1208 extern int explicit_null_checks_inserted,
1209            explicit_null_checks_elided;
1210 #endif
1211 Node* GraphKit::null_check_common(Node* value, BasicType type,
1212                                   // optional arguments for variations:
1213                                   bool assert_null,
1214                                   Node* *null_control,
1215                                   bool speculative) {
1216   assert(!assert_null || null_control == NULL, "not both at once");
1217   if (stopped())  return top();
1218   NOT_PRODUCT(explicit_null_checks_inserted++);
1219 
1220   // Construct NULL check
1221   Node *chk = NULL;
1222   switch(type) {
1223     case T_LONG   : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1224     case T_INT    : chk = new CmpINode(value, _gvn.intcon(0)); break;
1225     case T_VALUETYPE : // fall through
1226     case T_ARRAY  : // fall through
1227       type = T_OBJECT;  // simplify further tests
1228     case T_OBJECT : {
1229       const Type *t = _gvn.type( value );
1230 
1231       const TypeOopPtr* tp = t->isa_oopptr();
1232       if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1233           // Only for do_null_check, not any of its siblings:
1234           && !assert_null && null_control == NULL) {
1235         // Usually, any field access or invocation on an unloaded oop type
1236         // will simply fail to link, since the statically linked class is
1237         // likely also to be unloaded.  However, in -Xcomp mode, sometimes
1238         // the static class is loaded but the sharper oop type is not.
1239         // Rather than checking for this obscure case in lots of places,
1240         // we simply observe that a null check on an unloaded class
1241         // will always be followed by a nonsense operation, so we
1242         // can just issue the uncommon trap here.
1243         // Our access to the unloaded class will only be correct
1244         // after it has been loaded and initialized, which requires
1245         // a trip through the interpreter.
1246 #ifndef PRODUCT
1247         if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1248 #endif
1249         uncommon_trap(Deoptimization::Reason_unloaded,
1250                       Deoptimization::Action_reinterpret,
1251                       tp->klass(), "!loaded");
1252         return top();
1253       }
1254 
1255       if (assert_null) {
1256         // See if the type is contained in NULL_PTR.
1257         // If so, then the value is already null.
1258         if (t->higher_equal(TypePtr::NULL_PTR)) {
1259           NOT_PRODUCT(explicit_null_checks_elided++);
1260           return value;           // Elided null assert quickly!
1261         }
1262       } else {
1263         // See if mixing in the NULL pointer changes type.
1264         // If so, then the NULL pointer was not allowed in the original
1265         // type.  In other words, "value" was not-null.
1266         if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1267           // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1268           NOT_PRODUCT(explicit_null_checks_elided++);
1269           return value;           // Elided null check quickly!
1270         }
1271       }
1272       chk = new CmpPNode( value, null() );
1273       break;
1274     }
1275 
1276     default:
1277       fatal("unexpected type: %s", type2name(type));
1278   }
1279   assert(chk != NULL, "sanity check");
1280   chk = _gvn.transform(chk);
1281 
1282   BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1283   BoolNode *btst = new BoolNode( chk, btest);
1284   Node   *tst = _gvn.transform( btst );
1285 
1286   //-----------
1287   // if peephole optimizations occurred, a prior test existed.
1288   // If a prior test existed, maybe it dominates as we can avoid this test.
1289   if (tst != btst && type == T_OBJECT) {
1290     // At this point we want to scan up the CFG to see if we can
1291     // find an identical test (and so avoid this test altogether).
1292     Node *cfg = control();
1293     int depth = 0;
1294     while( depth < 16 ) {       // Limit search depth for speed
1295       if( cfg->Opcode() == Op_IfTrue &&
1296           cfg->in(0)->in(1) == tst ) {
1297         // Found prior test.  Use "cast_not_null" to construct an identical
1298         // CastPP (and hence hash to) as already exists for the prior test.
1299         // Return that casted value.
1300         if (assert_null) {
1301           replace_in_map(value, null());
1302           return null();  // do not issue the redundant test
1303         }
1304         Node *oldcontrol = control();
1305         set_control(cfg);
1306         Node *res = cast_not_null(value);
1307         set_control(oldcontrol);
1308         NOT_PRODUCT(explicit_null_checks_elided++);
1309         return res;
1310       }
1311       cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1312       if (cfg == NULL)  break;  // Quit at region nodes
1313       depth++;
1314     }
1315   }
1316 
1317   //-----------
1318   // Branch to failure if null
1319   float ok_prob = PROB_MAX;  // a priori estimate:  nulls never happen
1320   Deoptimization::DeoptReason reason;
1321   if (assert_null) {
1322     reason = Deoptimization::reason_null_assert(speculative);
1323   } else if (type == T_OBJECT) {
1324     reason = Deoptimization::reason_null_check(speculative);
1325   } else {
1326     reason = Deoptimization::Reason_div0_check;
1327   }
1328   // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1329   // ciMethodData::has_trap_at will return a conservative -1 if any
1330   // must-be-null assertion has failed.  This could cause performance
1331   // problems for a method after its first do_null_assert failure.
1332   // Consider using 'Reason_class_check' instead?
1333 
1334   // To cause an implicit null check, we set the not-null probability
1335   // to the maximum (PROB_MAX).  For an explicit check the probability
1336   // is set to a smaller value.
1337   if (null_control != NULL || too_many_traps(reason)) {
1338     // probability is less likely
1339     ok_prob =  PROB_LIKELY_MAG(3);
1340   } else if (!assert_null &&
1341              (ImplicitNullCheckThreshold > 0) &&
1342              method() != NULL &&
1343              (method()->method_data()->trap_count(reason)
1344               >= (uint)ImplicitNullCheckThreshold)) {
1345     ok_prob =  PROB_LIKELY_MAG(3);
1346   }
1347 
1348   if (null_control != NULL) {
1349     IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1350     Node* null_true = _gvn.transform( new IfFalseNode(iff));
1351     set_control(      _gvn.transform( new IfTrueNode(iff)));
1352 #ifndef PRODUCT
1353     if (null_true == top()) {
1354       explicit_null_checks_elided++;
1355     }
1356 #endif
1357     (*null_control) = null_true;
1358   } else {
1359     BuildCutout unless(this, tst, ok_prob);
1360     // Check for optimizer eliding test at parse time
1361     if (stopped()) {
1362       // Failure not possible; do not bother making uncommon trap.
1363       NOT_PRODUCT(explicit_null_checks_elided++);
1364     } else if (assert_null) {
1365       uncommon_trap(reason,
1366                     Deoptimization::Action_make_not_entrant,
1367                     NULL, "assert_null");
1368     } else {
1369       replace_in_map(value, zerocon(type));
1370       builtin_throw(reason);
1371     }
1372   }
1373 
1374   // Must throw exception, fall-thru not possible?
1375   if (stopped()) {
1376     return top();               // No result
1377   }
1378 
1379   if (assert_null) {
1380     // Cast obj to null on this path.
1381     replace_in_map(value, zerocon(type));
1382     return zerocon(type);
1383   }
1384 
1385   // Cast obj to not-null on this path, if there is no null_control.
1386   // (If there is a null_control, a non-null value may come back to haunt us.)
1387   if (type == T_OBJECT) {
1388     Node* cast = cast_not_null(value, false);
1389     if (null_control == NULL || (*null_control) == top())
1390       replace_in_map(value, cast);
1391     value = cast;
1392   }
1393 
1394   return value;
1395 }
1396 
1397 
1398 //------------------------------cast_not_null----------------------------------
1399 // Cast obj to not-null on this path
1400 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1401   assert(!obj->is_ValueType(), "should not cast value type");
1402   const Type *t = _gvn.type(obj);
1403   const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1404   // Object is already not-null?
1405   if( t == t_not_null ) return obj;
1406 
1407   Node *cast = new CastPPNode(obj,t_not_null);
1408   cast->init_req(0, control());
1409   cast = _gvn.transform( cast );
1410 
1411   // Scan for instances of 'obj' in the current JVM mapping.
1412   // These instances are known to be not-null after the test.
1413   if (do_replace_in_map)
1414     replace_in_map(obj, cast);
1415 
1416   return cast;                  // Return casted value
1417 }
1418 
1419 // Sometimes in intrinsics, we implicitly know an object is not null
1420 // (there's no actual null check) so we can cast it to not null. In
1421 // the course of optimizations, the input to the cast can become null.
1422 // In that case that data path will die and we need the control path
1423 // to become dead as well to keep the graph consistent. So we have to
1424 // add a check for null for which one branch can't be taken. It uses
1425 // an Opaque4 node that will cause the check to be removed after loop
1426 // opts so the test goes away and the compiled code doesn't execute a
1427 // useless check.
1428 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1429   Node* chk = _gvn.transform(new CmpPNode(value, null()));
1430   Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1431   Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1)));
1432   IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1433   _gvn.set_type(iff, iff->Value(&_gvn));
1434   Node *if_f = _gvn.transform(new IfFalseNode(iff));
1435   Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr));
1436   Node *halt = _gvn.transform(new HaltNode(if_f, frame));
1437   C->root()->add_req(halt);
1438   Node *if_t = _gvn.transform(new IfTrueNode(iff));
1439   set_control(if_t);
1440   return cast_not_null(value, do_replace_in_map);
1441 }
1442 
1443 
1444 //--------------------------replace_in_map-------------------------------------
1445 void GraphKit::replace_in_map(Node* old, Node* neww) {
1446   if (old == neww) {
1447     return;
1448   }
1449 
1450   map()->replace_edge(old, neww);
1451 
1452   // Note: This operation potentially replaces any edge
1453   // on the map.  This includes locals, stack, and monitors
1454   // of the current (innermost) JVM state.
1455 
1456   // don't let inconsistent types from profiling escape this
1457   // method
1458 
1459   const Type* told = _gvn.type(old);
1460   const Type* tnew = _gvn.type(neww);
1461 
1462   if (!tnew->higher_equal(told)) {
1463     return;
1464   }
1465 
1466   map()->record_replaced_node(old, neww);
1467 }
1468 
1469 
1470 //=============================================================================
1471 //--------------------------------memory---------------------------------------
1472 Node* GraphKit::memory(uint alias_idx) {
1473   MergeMemNode* mem = merged_memory();
1474   Node* p = mem->memory_at(alias_idx);
1475   _gvn.set_type(p, Type::MEMORY);  // must be mapped
1476   return p;
1477 }
1478 
1479 //-----------------------------reset_memory------------------------------------
1480 Node* GraphKit::reset_memory() {
1481   Node* mem = map()->memory();
1482   // do not use this node for any more parsing!
1483   debug_only( map()->set_memory((Node*)NULL) );
1484   return _gvn.transform( mem );
1485 }
1486 
1487 //------------------------------set_all_memory---------------------------------
1488 void GraphKit::set_all_memory(Node* newmem) {
1489   Node* mergemem = MergeMemNode::make(newmem);
1490   gvn().set_type_bottom(mergemem);
1491   map()->set_memory(mergemem);
1492 }
1493 
1494 //------------------------------set_all_memory_call----------------------------
1495 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1496   Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1497   set_all_memory(newmem);
1498 }
1499 
1500 //=============================================================================
1501 //
1502 // parser factory methods for MemNodes
1503 //
1504 // These are layered on top of the factory methods in LoadNode and StoreNode,
1505 // and integrate with the parser's memory state and _gvn engine.
1506 //
1507 
1508 // factory methods in "int adr_idx"
1509 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1510                           int adr_idx,
1511                           MemNode::MemOrd mo,
1512                           LoadNode::ControlDependency control_dependency,
1513                           bool require_atomic_access,
1514                           bool unaligned,
1515                           bool mismatched) {
1516   assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1517   const TypePtr* adr_type = NULL; // debug-mode-only argument
1518   debug_only(adr_type = C->get_adr_type(adr_idx));
1519   Node* mem = memory(adr_idx);
1520   Node* ld;
1521   if (require_atomic_access && bt == T_LONG) {
1522     ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched);
1523   } else if (require_atomic_access && bt == T_DOUBLE) {
1524     ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched);
1525   } else {
1526     ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched);
1527   }
1528   ld = _gvn.transform(ld);
1529 
1530   if (((bt == T_OBJECT || bt == T_VALUETYPE || bt == T_VALUETYPEPTR) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1531     // Improve graph before escape analysis and boxing elimination.
1532     record_for_igvn(ld);
1533   }
1534   return ld;
1535 }
1536 
1537 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1538                                 int adr_idx,
1539                                 MemNode::MemOrd mo,
1540                                 bool require_atomic_access,
1541                                 bool unaligned,
1542                                 bool mismatched) {
1543   assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1544   const TypePtr* adr_type = NULL;
1545   debug_only(adr_type = C->get_adr_type(adr_idx));
1546   Node *mem = memory(adr_idx);
1547   Node* st;
1548   if (require_atomic_access && bt == T_LONG) {
1549     st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1550   } else if (require_atomic_access && bt == T_DOUBLE) {
1551     st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1552   } else {
1553     st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo);
1554   }
1555   if (unaligned) {
1556     st->as_Store()->set_unaligned_access();
1557   }
1558   if (mismatched) {
1559     st->as_Store()->set_mismatched_access();
1560   }
1561   st = _gvn.transform(st);
1562   set_memory(st, adr_idx);
1563   // Back-to-back stores can only remove intermediate store with DU info
1564   // so push on worklist for optimizer.
1565   if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1566     record_for_igvn(st);
1567 
1568   return st;
1569 }
1570 
1571 Node* GraphKit::access_store_at(Node* ctl,
1572                                 Node* obj,
1573                                 Node* adr,
1574                                 const TypePtr* adr_type,
1575                                 Node* val,
1576                                 const Type* val_type,
1577                                 BasicType bt,
1578                                 DecoratorSet decorators,
1579                                 bool deoptimize_on_exception) {
1580   // Transformation of a value which could be NULL pointer (CastPP #NULL)
1581   // could be delayed during Parse (for example, in adjust_map_after_if()).
1582   // Execute transformation here to avoid barrier generation in such case.
1583   if (_gvn.type(val) == TypePtr::NULL_PTR) {
1584     val = _gvn.makecon(TypePtr::NULL_PTR);
1585   }
1586 
1587   set_control(ctl);
1588   if (stopped()) {
1589     return top(); // Dead path ?
1590   }
1591 
1592   assert(val != NULL, "not dead path");
1593   if (val->is_ValueType()) {
1594     // Allocate value type and get oop
1595     val = val->as_ValueType()->allocate(this, deoptimize_on_exception)->get_oop();
1596   }
1597 
1598   C2AccessValuePtr addr(adr, adr_type);
1599   C2AccessValue value(val, val_type);
1600   C2Access access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr);
1601   if (access.is_raw()) {
1602     return _barrier_set->BarrierSetC2::store_at(access, value);
1603   } else {
1604     return _barrier_set->store_at(access, value);
1605   }
1606 }
1607 
1608 Node* GraphKit::access_load_at(Node* obj,   // containing obj
1609                                Node* adr,   // actual adress to store val at
1610                                const TypePtr* adr_type,
1611                                const Type* val_type,
1612                                BasicType bt,
1613                                DecoratorSet decorators) {
1614   if (stopped()) {
1615     return top(); // Dead path ?
1616   }
1617 
1618   C2AccessValuePtr addr(adr, adr_type);
1619   C2Access access(this, decorators | C2_READ_ACCESS, bt, obj, addr);
1620   if (access.is_raw()) {
1621     return _barrier_set->BarrierSetC2::load_at(access, val_type);
1622   } else {
1623     return _barrier_set->load_at(access, val_type);
1624   }
1625 }
1626 
1627 Node* GraphKit::access_load(Node* adr,   // actual adress to load val at
1628                             const Type* val_type,
1629                             BasicType bt,
1630                             DecoratorSet decorators) {
1631   if (stopped()) {
1632     return top(); // Dead path ?
1633   }
1634 
1635   C2AccessValuePtr addr(adr, NULL);
1636   C2Access access(this, decorators | C2_READ_ACCESS, bt, NULL, addr);
1637   if (access.is_raw()) {
1638     return _barrier_set->BarrierSetC2::load_at(access, val_type);
1639   } else {
1640     return _barrier_set->load_at(access, val_type);
1641   }
1642 }
1643 
1644 Node* GraphKit::access_atomic_cmpxchg_val_at(Node* ctl,
1645                                              Node* obj,
1646                                              Node* adr,
1647                                              const TypePtr* adr_type,
1648                                              int alias_idx,
1649                                              Node* expected_val,
1650                                              Node* new_val,
1651                                              const Type* value_type,
1652                                              BasicType bt,
1653                                              DecoratorSet decorators) {
1654   set_control(ctl);
1655   C2AccessValuePtr addr(adr, adr_type);
1656   C2AtomicAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1657                         bt, obj, addr, alias_idx);
1658   if (access.is_raw()) {
1659     return _barrier_set->BarrierSetC2::atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1660   } else {
1661     return _barrier_set->atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1662   }
1663 }
1664 
1665 Node* GraphKit::access_atomic_cmpxchg_bool_at(Node* ctl,
1666                                               Node* obj,
1667                                               Node* adr,
1668                                               const TypePtr* adr_type,
1669                                               int alias_idx,
1670                                               Node* expected_val,
1671                                               Node* new_val,
1672                                               const Type* value_type,
1673                                               BasicType bt,
1674                                               DecoratorSet decorators) {
1675   set_control(ctl);
1676   C2AccessValuePtr addr(adr, adr_type);
1677   C2AtomicAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1678                         bt, obj, addr, alias_idx);
1679   if (access.is_raw()) {
1680     return _barrier_set->BarrierSetC2::atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1681   } else {
1682     return _barrier_set->atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1683   }
1684 }
1685 
1686 Node* GraphKit::access_atomic_xchg_at(Node* ctl,
1687                                       Node* obj,
1688                                       Node* adr,
1689                                       const TypePtr* adr_type,
1690                                       int alias_idx,
1691                                       Node* new_val,
1692                                       const Type* value_type,
1693                                       BasicType bt,
1694                                       DecoratorSet decorators) {
1695   set_control(ctl);
1696   C2AccessValuePtr addr(adr, adr_type);
1697   C2AtomicAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1698                         bt, obj, addr, alias_idx);
1699   if (access.is_raw()) {
1700     return _barrier_set->BarrierSetC2::atomic_xchg_at(access, new_val, value_type);
1701   } else {
1702     return _barrier_set->atomic_xchg_at(access, new_val, value_type);
1703   }
1704 }
1705 
1706 Node* GraphKit::access_atomic_add_at(Node* ctl,
1707                                      Node* obj,
1708                                      Node* adr,
1709                                      const TypePtr* adr_type,
1710                                      int alias_idx,
1711                                      Node* new_val,
1712                                      const Type* value_type,
1713                                      BasicType bt,
1714                                      DecoratorSet decorators) {
1715   set_control(ctl);
1716   C2AccessValuePtr addr(adr, adr_type);
1717   C2AtomicAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1718   if (access.is_raw()) {
1719     return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1720   } else {
1721     return _barrier_set->atomic_add_at(access, new_val, value_type);
1722   }
1723 }
1724 
1725 void GraphKit::access_clone(Node* ctl, Node* src_base, Node* dst_base, Node* countx, bool is_array) {
1726   set_control(ctl);
1727   return _barrier_set->clone(this, src_base, dst_base, countx, is_array);
1728 }
1729 
1730 //-------------------------array_element_address-------------------------
1731 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1732                                       const TypeInt* sizetype, Node* ctrl) {
1733   uint shift  = exact_log2(type2aelembytes(elembt));
1734   ciKlass* arytype_klass = _gvn.type(ary)->is_aryptr()->klass();
1735   if (arytype_klass->is_value_array_klass()) {
1736     ciValueArrayKlass* vak = arytype_klass->as_value_array_klass();
1737     shift = vak->log2_element_size();
1738   }
1739   uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1740 
1741   // short-circuit a common case (saves lots of confusing waste motion)
1742   jint idx_con = find_int_con(idx, -1);
1743   if (idx_con >= 0) {
1744     intptr_t offset = header + ((intptr_t)idx_con << shift);
1745     return basic_plus_adr(ary, offset);
1746   }
1747 
1748   // must be correct type for alignment purposes
1749   Node* base  = basic_plus_adr(ary, header);
1750   idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1751   Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1752   return basic_plus_adr(ary, base, scale);
1753 }
1754 
1755 //-------------------------load_array_element-------------------------
1756 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1757   const Type* elemtype = arytype->elem();
1758   BasicType elembt = elemtype->array_element_basic_type();
1759   assert(elembt != T_VALUETYPE, "value types are not supported by this method");
1760   Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1761   if (elembt == T_NARROWOOP) {
1762     elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1763   }
1764   Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered);
1765   return ld;
1766 }
1767 
1768 //-------------------------set_arguments_for_java_call-------------------------
1769 // Arguments (pre-popped from the stack) are taken from the JVMS.
1770 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1771   // Add the call arguments:
1772   const TypeTuple* domain = call->tf()->domain_sig();
1773   uint nargs = domain->cnt();
1774   for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) {
1775     Node* arg = argument(i-TypeFunc::Parms);
1776     const Type* t = domain->field_at(i);
1777     if (arg->is_ValueType()) {
1778       assert(t->is_oopptr()->can_be_value_type(), "wrong argument type");
1779       ValueTypeNode* vt = arg->as_ValueType();
1780       if (ValueTypePassFieldsAsArgs) {
1781         // We don't pass value type arguments by reference but instead
1782         // pass each field of the value type
1783         idx += vt->pass_fields(call, idx, *this);
1784         // If a value type argument is passed as fields, attach the Method* to the call site
1785         // to be able to access the extended signature later via attached_method_before_pc().
1786         // For example, see CompiledMethod::preserve_callee_argument_oops().
1787         call->set_override_symbolic_info(true);
1788         continue;
1789       } else {
1790         // Pass value type argument via oop to callee
1791         arg = vt->allocate(this)->get_oop();
1792       }
1793     } else if (t->is_valuetypeptr()) {
1794       ciMethod* declared_method = method()->get_method_at_bci(bci());
1795       if (arg->is_ValueTypePtr()) {
1796         // We are calling an Object method with a value type receiver
1797         ValueTypePtrNode* vt = arg->isa_ValueTypePtr();
1798         assert(i == TypeFunc::Parms && !declared_method->is_static(), "argument must be receiver");
1799         assert(vt->is_allocated(&gvn()), "argument must be allocated");
1800         arg = vt->get_oop();
1801       } else {
1802         // Constant null passed for a value type argument
1803         assert(arg->bottom_type()->remove_speculative() == TypePtr::NULL_PTR, "Anything other than null?");
1804         int arg_size = declared_method->signature()->arg_size_for_bc(java_bc());
1805         inc_sp(arg_size); // restore arguments
1806         uncommon_trap(Deoptimization::Reason_null_check, Deoptimization::Action_none);
1807         return;
1808       }
1809     }
1810     call->init_req(idx, arg);
1811     idx++;
1812   }
1813 }
1814 
1815 //---------------------------set_edges_for_java_call---------------------------
1816 // Connect a newly created call into the current JVMS.
1817 // A return value node (if any) is returned from set_edges_for_java_call.
1818 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1819 
1820   // Add the predefined inputs:
1821   call->init_req( TypeFunc::Control, control() );
1822   call->init_req( TypeFunc::I_O    , i_o() );
1823   call->init_req( TypeFunc::Memory , reset_memory() );
1824   call->init_req( TypeFunc::FramePtr, frameptr() );
1825   call->init_req( TypeFunc::ReturnAdr, top() );
1826 
1827   add_safepoint_edges(call, must_throw);
1828 
1829   Node* xcall = _gvn.transform(call);
1830 
1831   if (xcall == top()) {
1832     set_control(top());
1833     return;
1834   }
1835   assert(xcall == call, "call identity is stable");
1836 
1837   // Re-use the current map to produce the result.
1838 
1839   set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1840   set_i_o(    _gvn.transform(new ProjNode(call, TypeFunc::I_O    , separate_io_proj)));
1841   set_all_memory_call(xcall, separate_io_proj);
1842 
1843   //return xcall;   // no need, caller already has it
1844 }
1845 
1846 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
1847   if (stopped())  return top();  // maybe the call folded up?
1848 
1849   // Note:  Since any out-of-line call can produce an exception,
1850   // we always insert an I_O projection from the call into the result.
1851 
1852   make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
1853 
1854   if (separate_io_proj) {
1855     // The caller requested separate projections be used by the fall
1856     // through and exceptional paths, so replace the projections for
1857     // the fall through path.
1858     set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1859     set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1860   }
1861 
1862   // Capture the return value, if any.
1863   Node* ret;
1864   if (call->method() == NULL ||
1865       call->method()->return_type()->basic_type() == T_VOID) {
1866     ret = top();
1867   } else {
1868     if (!call->tf()->returns_value_type_as_fields()) {
1869       ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1870     } else {
1871       // Return of multiple values (value type fields): we create a
1872       // ValueType node, each field is a projection from the call.
1873       const TypeTuple* range_sig = call->tf()->range_sig();
1874       const Type* t = range_sig->field_at(TypeFunc::Parms);
1875       assert(t->is_valuetypeptr(), "only value types for multiple return values");
1876       ciValueKlass* vk = t->value_klass();
1877       ret = ValueTypeNode::make_from_multi(this, call, vk, TypeFunc::Parms+1, false);
1878     }
1879   }
1880 
1881   return ret;
1882 }
1883 
1884 //--------------------set_predefined_input_for_runtime_call--------------------
1885 // Reading and setting the memory state is way conservative here.
1886 // The real problem is that I am not doing real Type analysis on memory,
1887 // so I cannot distinguish card mark stores from other stores.  Across a GC
1888 // point the Store Barrier and the card mark memory has to agree.  I cannot
1889 // have a card mark store and its barrier split across the GC point from
1890 // either above or below.  Here I get that to happen by reading ALL of memory.
1891 // A better answer would be to separate out card marks from other memory.
1892 // For now, return the input memory state, so that it can be reused
1893 // after the call, if this call has restricted memory effects.
1894 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1895   // Set fixed predefined input arguments
1896   Node* memory = reset_memory();
1897   call->init_req( TypeFunc::Control,   control()  );
1898   call->init_req( TypeFunc::I_O,       top()      ); // does no i/o
1899   call->init_req( TypeFunc::Memory,    memory     ); // may gc ptrs
1900   call->init_req( TypeFunc::FramePtr,  frameptr() );
1901   call->init_req( TypeFunc::ReturnAdr, top()      );
1902   return memory;
1903 }
1904 
1905 //-------------------set_predefined_output_for_runtime_call--------------------
1906 // Set control and memory (not i_o) from the call.
1907 // If keep_mem is not NULL, use it for the output state,
1908 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1909 // If hook_mem is NULL, this call produces no memory effects at all.
1910 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1911 // then only that memory slice is taken from the call.
1912 // In the last case, we must put an appropriate memory barrier before
1913 // the call, so as to create the correct anti-dependencies on loads
1914 // preceding the call.
1915 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1916                                                       Node* keep_mem,
1917                                                       const TypePtr* hook_mem) {
1918   // no i/o
1919   set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) ));
1920   if (keep_mem) {
1921     // First clone the existing memory state
1922     set_all_memory(keep_mem);
1923     if (hook_mem != NULL) {
1924       // Make memory for the call
1925       Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) );
1926       // Set the RawPtr memory state only.  This covers all the heap top/GC stuff
1927       // We also use hook_mem to extract specific effects from arraycopy stubs.
1928       set_memory(mem, hook_mem);
1929     }
1930     // ...else the call has NO memory effects.
1931 
1932     // Make sure the call advertises its memory effects precisely.
1933     // This lets us build accurate anti-dependences in gcm.cpp.
1934     assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1935            "call node must be constructed correctly");
1936   } else {
1937     assert(hook_mem == NULL, "");
1938     // This is not a "slow path" call; all memory comes from the call.
1939     set_all_memory_call(call);
1940   }
1941 }
1942 
1943 
1944 // Replace the call with the current state of the kit.
1945 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
1946   JVMState* ejvms = NULL;
1947   if (has_exceptions()) {
1948     ejvms = transfer_exceptions_into_jvms();
1949   }
1950 
1951   ReplacedNodes replaced_nodes = map()->replaced_nodes();
1952   ReplacedNodes replaced_nodes_exception;
1953   Node* ex_ctl = top();
1954 
1955   SafePointNode* final_state = stop();
1956 
1957   // Find all the needed outputs of this call
1958   CallProjections* callprojs = call->extract_projections(true);
1959 
1960   Node* init_mem = call->in(TypeFunc::Memory);
1961   Node* final_mem = final_state->in(TypeFunc::Memory);
1962   Node* final_ctl = final_state->in(TypeFunc::Control);
1963   Node* final_io = final_state->in(TypeFunc::I_O);
1964 
1965   // Replace all the old call edges with the edges from the inlining result
1966   if (callprojs->fallthrough_catchproj != NULL) {
1967     C->gvn_replace_by(callprojs->fallthrough_catchproj, final_ctl);
1968   }
1969   if (callprojs->fallthrough_memproj != NULL) {
1970     if (final_mem->is_MergeMem()) {
1971       // Parser's exits MergeMem was not transformed but may be optimized
1972       final_mem = _gvn.transform(final_mem);
1973     }
1974     C->gvn_replace_by(callprojs->fallthrough_memproj,   final_mem);
1975   }
1976   if (callprojs->fallthrough_ioproj != NULL) {
1977     C->gvn_replace_by(callprojs->fallthrough_ioproj,    final_io);
1978   }
1979 
1980   // Replace the result with the new result if it exists and is used
1981   if (callprojs->resproj[0] != NULL && result != NULL) {
1982     assert(callprojs->nb_resproj == 1, "unexpected number of results");
1983     C->gvn_replace_by(callprojs->resproj[0], result);
1984   }
1985 
1986   if (ejvms == NULL) {
1987     // No exception edges to simply kill off those paths
1988     if (callprojs->catchall_catchproj != NULL) {
1989       C->gvn_replace_by(callprojs->catchall_catchproj, C->top());
1990     }
1991     if (callprojs->catchall_memproj != NULL) {
1992       C->gvn_replace_by(callprojs->catchall_memproj,   C->top());
1993     }
1994     if (callprojs->catchall_ioproj != NULL) {
1995       C->gvn_replace_by(callprojs->catchall_ioproj,    C->top());
1996     }
1997     // Replace the old exception object with top
1998     if (callprojs->exobj != NULL) {
1999       C->gvn_replace_by(callprojs->exobj, C->top());
2000     }
2001   } else {
2002     GraphKit ekit(ejvms);
2003 
2004     // Load my combined exception state into the kit, with all phis transformed:
2005     SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
2006     replaced_nodes_exception = ex_map->replaced_nodes();
2007 
2008     Node* ex_oop = ekit.use_exception_state(ex_map);
2009 
2010     if (callprojs->catchall_catchproj != NULL) {
2011       C->gvn_replace_by(callprojs->catchall_catchproj, ekit.control());
2012       ex_ctl = ekit.control();
2013     }
2014     if (callprojs->catchall_memproj != NULL) {
2015       C->gvn_replace_by(callprojs->catchall_memproj,   ekit.reset_memory());
2016     }
2017     if (callprojs->catchall_ioproj != NULL) {
2018       C->gvn_replace_by(callprojs->catchall_ioproj,    ekit.i_o());
2019     }
2020 
2021     // Replace the old exception object with the newly created one
2022     if (callprojs->exobj != NULL) {
2023       C->gvn_replace_by(callprojs->exobj, ex_oop);
2024     }
2025   }
2026 
2027   // Disconnect the call from the graph
2028   call->disconnect_inputs(NULL, C);
2029   C->gvn_replace_by(call, C->top());
2030 
2031   // Clean up any MergeMems that feed other MergeMems since the
2032   // optimizer doesn't like that.
2033   if (final_mem->is_MergeMem()) {
2034     Node_List wl;
2035     for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) {
2036       Node* m = i.get();
2037       if (m->is_MergeMem() && !wl.contains(m)) {
2038         wl.push(m);
2039       }
2040     }
2041     while (wl.size()  > 0) {
2042       _gvn.transform(wl.pop());
2043     }
2044   }
2045 
2046   if (callprojs->fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) {
2047     replaced_nodes.apply(C, final_ctl);
2048   }
2049   if (!ex_ctl->is_top() && do_replaced_nodes) {
2050     replaced_nodes_exception.apply(C, ex_ctl);
2051   }
2052 }
2053 
2054 
2055 //------------------------------increment_counter------------------------------
2056 // for statistics: increment a VM counter by 1
2057 
2058 void GraphKit::increment_counter(address counter_addr) {
2059   Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2060   increment_counter(adr1);
2061 }
2062 
2063 void GraphKit::increment_counter(Node* counter_addr) {
2064   int adr_type = Compile::AliasIdxRaw;
2065   Node* ctrl = control();
2066   Node* cnt  = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
2067   Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1)));
2068   store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered);
2069 }
2070 
2071 
2072 //------------------------------uncommon_trap----------------------------------
2073 // Bail out to the interpreter in mid-method.  Implemented by calling the
2074 // uncommon_trap blob.  This helper function inserts a runtime call with the
2075 // right debug info.
2076 void GraphKit::uncommon_trap(int trap_request,
2077                              ciKlass* klass, const char* comment,
2078                              bool must_throw,
2079                              bool keep_exact_action) {
2080   if (failing())  stop();
2081   if (stopped())  return; // trap reachable?
2082 
2083   // Note:  If ProfileTraps is true, and if a deopt. actually
2084   // occurs here, the runtime will make sure an MDO exists.  There is
2085   // no need to call method()->ensure_method_data() at this point.
2086 
2087   // Set the stack pointer to the right value for reexecution:
2088   set_sp(reexecute_sp());
2089 
2090 #ifdef ASSERT
2091   if (!must_throw) {
2092     // Make sure the stack has at least enough depth to execute
2093     // the current bytecode.
2094     int inputs, ignored_depth;
2095     if (compute_stack_effects(inputs, ignored_depth)) {
2096       assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
2097              Bytecodes::name(java_bc()), sp(), inputs);
2098     }
2099   }
2100 #endif
2101 
2102   Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
2103   Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
2104 
2105   switch (action) {
2106   case Deoptimization::Action_maybe_recompile:
2107   case Deoptimization::Action_reinterpret:
2108     // Temporary fix for 6529811 to allow virtual calls to be sure they
2109     // get the chance to go from mono->bi->mega
2110     if (!keep_exact_action &&
2111         Deoptimization::trap_request_index(trap_request) < 0 &&
2112         too_many_recompiles(reason)) {
2113       // This BCI is causing too many recompilations.
2114       if (C->log() != NULL) {
2115         C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
2116                 Deoptimization::trap_reason_name(reason),
2117                 Deoptimization::trap_action_name(action));
2118       }
2119       action = Deoptimization::Action_none;
2120       trap_request = Deoptimization::make_trap_request(reason, action);
2121     } else {
2122       C->set_trap_can_recompile(true);
2123     }
2124     break;
2125   case Deoptimization::Action_make_not_entrant:
2126     C->set_trap_can_recompile(true);
2127     break;
2128   case Deoptimization::Action_none:
2129   case Deoptimization::Action_make_not_compilable:
2130     break;
2131   default:
2132 #ifdef ASSERT
2133     fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action));
2134 #endif
2135     break;
2136   }
2137 
2138   if (TraceOptoParse) {
2139     char buf[100];
2140     tty->print_cr("Uncommon trap %s at bci:%d",
2141                   Deoptimization::format_trap_request(buf, sizeof(buf),
2142                                                       trap_request), bci());
2143   }
2144 
2145   CompileLog* log = C->log();
2146   if (log != NULL) {
2147     int kid = (klass == NULL)? -1: log->identify(klass);
2148     log->begin_elem("uncommon_trap bci='%d'", bci());
2149     char buf[100];
2150     log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2151                                                           trap_request));
2152     if (kid >= 0)         log->print(" klass='%d'", kid);
2153     if (comment != NULL)  log->print(" comment='%s'", comment);
2154     log->end_elem();
2155   }
2156 
2157   // Make sure any guarding test views this path as very unlikely
2158   Node *i0 = control()->in(0);
2159   if (i0 != NULL && i0->is_If()) {        // Found a guarding if test?
2160     IfNode *iff = i0->as_If();
2161     float f = iff->_prob;   // Get prob
2162     if (control()->Opcode() == Op_IfTrue) {
2163       if (f > PROB_UNLIKELY_MAG(4))
2164         iff->_prob = PROB_MIN;
2165     } else {
2166       if (f < PROB_LIKELY_MAG(4))
2167         iff->_prob = PROB_MAX;
2168     }
2169   }
2170 
2171   // Clear out dead values from the debug info.
2172   kill_dead_locals();
2173 
2174   // Now insert the uncommon trap subroutine call
2175   address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2176   const TypePtr* no_memory_effects = NULL;
2177   // Pass the index of the class to be loaded
2178   Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2179                                  (must_throw ? RC_MUST_THROW : 0),
2180                                  OptoRuntime::uncommon_trap_Type(),
2181                                  call_addr, "uncommon_trap", no_memory_effects,
2182                                  intcon(trap_request));
2183   assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2184          "must extract request correctly from the graph");
2185   assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2186 
2187   call->set_req(TypeFunc::ReturnAdr, returnadr());
2188   // The debug info is the only real input to this call.
2189 
2190   // Halt-and-catch fire here.  The above call should never return!
2191   HaltNode* halt = new HaltNode(control(), frameptr());
2192   _gvn.set_type_bottom(halt);
2193   root()->add_req(halt);
2194 
2195   stop_and_kill_map();
2196 }
2197 
2198 
2199 //--------------------------just_allocated_object------------------------------
2200 // Report the object that was just allocated.
2201 // It must be the case that there are no intervening safepoints.
2202 // We use this to determine if an object is so "fresh" that
2203 // it does not require card marks.
2204 Node* GraphKit::just_allocated_object(Node* current_control) {
2205   if (C->recent_alloc_ctl() == current_control)
2206     return C->recent_alloc_obj();
2207   return NULL;
2208 }
2209 
2210 
2211 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2212   // (Note:  TypeFunc::make has a cache that makes this fast.)
2213   const TypeFunc* tf    = TypeFunc::make(dest_method);
2214   int             nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2215   for (int j = 0; j < nargs; j++) {
2216     const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2217     if( targ->basic_type() == T_DOUBLE ) {
2218       // If any parameters are doubles, they must be rounded before
2219       // the call, dstore_rounding does gvn.transform
2220       Node *arg = argument(j);
2221       arg = dstore_rounding(arg);
2222       set_argument(j, arg);
2223     }
2224   }
2225 }
2226 
2227 /**
2228  * Record profiling data exact_kls for Node n with the type system so
2229  * that it can propagate it (speculation)
2230  *
2231  * @param n          node that the type applies to
2232  * @param exact_kls  type from profiling
2233  * @param maybe_null did profiling see null?
2234  *
2235  * @return           node with improved type
2236  */
2237 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2238   const Type* current_type = _gvn.type(n);
2239   assert(UseTypeSpeculation, "type speculation must be on");
2240 
2241   const TypePtr* speculative = current_type->speculative();
2242 
2243   // Should the klass from the profile be recorded in the speculative type?
2244   if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2245     const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls);
2246     const TypeOopPtr* xtype = tklass->as_instance_type();
2247     assert(xtype->klass_is_exact(), "Should be exact");
2248     // Any reason to believe n is not null (from this profiling or a previous one)?
2249     assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2250     const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2251     // record the new speculative type's depth
2252     speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2253     speculative = speculative->with_inline_depth(jvms()->depth());
2254   } else if (current_type->would_improve_ptr(ptr_kind)) {
2255     // Profiling report that null was never seen so we can change the
2256     // speculative type to non null ptr.
2257     if (ptr_kind == ProfileAlwaysNull) {
2258       speculative = TypePtr::NULL_PTR;
2259     } else {
2260       assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2261       const TypePtr* ptr = TypePtr::NOTNULL;
2262       if (speculative != NULL) {
2263         speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2264       } else {
2265         speculative = ptr;
2266       }
2267     }
2268   }
2269 
2270   if (speculative != current_type->speculative()) {
2271     // Build a type with a speculative type (what we think we know
2272     // about the type but will need a guard when we use it)
2273     const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative);
2274     // We're changing the type, we need a new CheckCast node to carry
2275     // the new type. The new type depends on the control: what
2276     // profiling tells us is only valid from here as far as we can
2277     // tell.
2278     Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2279     cast = _gvn.transform(cast);
2280     replace_in_map(n, cast);
2281     n = cast;
2282   }
2283 
2284   return n;
2285 }
2286 
2287 /**
2288  * Record profiling data from receiver profiling at an invoke with the
2289  * type system so that it can propagate it (speculation)
2290  *
2291  * @param n  receiver node
2292  *
2293  * @return   node with improved type
2294  */
2295 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2296   if (!UseTypeSpeculation) {
2297     return n;
2298   }
2299   ciKlass* exact_kls = profile_has_unique_klass();
2300   ProfilePtrKind ptr_kind = ProfileMaybeNull;
2301   if ((java_bc() == Bytecodes::_checkcast ||
2302        java_bc() == Bytecodes::_instanceof ||
2303        java_bc() == Bytecodes::_aastore) &&
2304       method()->method_data()->is_mature()) {
2305     ciProfileData* data = method()->method_data()->bci_to_data(bci());
2306     if (data != NULL) {
2307       if (!data->as_BitData()->null_seen()) {
2308         ptr_kind = ProfileNeverNull;
2309       } else {
2310         assert(data->is_ReceiverTypeData(), "bad profile data type");
2311         ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2312         uint i = 0;
2313         for (; i < call->row_limit(); i++) {
2314           ciKlass* receiver = call->receiver(i);
2315           if (receiver != NULL) {
2316             break;
2317           }
2318         }
2319         ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2320       }
2321     }
2322   }
2323   return record_profile_for_speculation(n, exact_kls, ptr_kind);
2324 }
2325 
2326 /**
2327  * Record profiling data from argument profiling at an invoke with the
2328  * type system so that it can propagate it (speculation)
2329  *
2330  * @param dest_method  target method for the call
2331  * @param bc           what invoke bytecode is this?
2332  */
2333 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2334   if (!UseTypeSpeculation) {
2335     return;
2336   }
2337   const TypeFunc* tf    = TypeFunc::make(dest_method);
2338   int             nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2339   int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2340   for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2341     const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2342     if (targ->isa_oopptr()) {
2343       ProfilePtrKind ptr_kind = ProfileMaybeNull;
2344       ciKlass* better_type = NULL;
2345       if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2346         record_profile_for_speculation(argument(j), better_type, ptr_kind);
2347       }
2348       i++;
2349     }
2350   }
2351 }
2352 
2353 /**
2354  * Record profiling data from parameter profiling at an invoke with
2355  * the type system so that it can propagate it (speculation)
2356  */
2357 void GraphKit::record_profiled_parameters_for_speculation() {
2358   if (!UseTypeSpeculation) {
2359     return;
2360   }
2361   for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2362     if (_gvn.type(local(i))->isa_oopptr()) {
2363       ProfilePtrKind ptr_kind = ProfileMaybeNull;
2364       ciKlass* better_type = NULL;
2365       if (method()->parameter_profiled_type(j, better_type, ptr_kind)) {
2366         record_profile_for_speculation(local(i), better_type, ptr_kind);
2367       }
2368       j++;
2369     }
2370   }
2371 }
2372 
2373 /**
2374  * Record profiling data from return value profiling at an invoke with
2375  * the type system so that it can propagate it (speculation)
2376  */
2377 void GraphKit::record_profiled_return_for_speculation() {
2378   if (!UseTypeSpeculation) {
2379     return;
2380   }
2381   ProfilePtrKind ptr_kind = ProfileMaybeNull;
2382   ciKlass* better_type = NULL;
2383   if (method()->return_profiled_type(bci(), better_type, ptr_kind)) {
2384     // If profiling reports a single type for the return value,
2385     // feed it to the type system so it can propagate it as a
2386     // speculative type
2387     record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind);
2388   }
2389 }
2390 
2391 void GraphKit::round_double_result(ciMethod* dest_method) {
2392   // A non-strict method may return a double value which has an extended
2393   // exponent, but this must not be visible in a caller which is 'strict'
2394   // If a strict caller invokes a non-strict callee, round a double result
2395 
2396   BasicType result_type = dest_method->return_type()->basic_type();
2397   assert( method() != NULL, "must have caller context");
2398   if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2399     // Destination method's return value is on top of stack
2400     // dstore_rounding() does gvn.transform
2401     Node *result = pop_pair();
2402     result = dstore_rounding(result);
2403     push_pair(result);
2404   }
2405 }
2406 
2407 // rounding for strict float precision conformance
2408 Node* GraphKit::precision_rounding(Node* n) {
2409   return UseStrictFP && _method->flags().is_strict()
2410     && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
2411     ? _gvn.transform( new RoundFloatNode(0, n) )
2412     : n;
2413 }
2414 
2415 // rounding for strict double precision conformance
2416 Node* GraphKit::dprecision_rounding(Node *n) {
2417   return UseStrictFP && _method->flags().is_strict()
2418     && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
2419     ? _gvn.transform( new RoundDoubleNode(0, n) )
2420     : n;
2421 }
2422 
2423 // rounding for non-strict double stores
2424 Node* GraphKit::dstore_rounding(Node* n) {
2425   return Matcher::strict_fp_requires_explicit_rounding
2426     && UseSSE <= 1
2427     ? _gvn.transform( new RoundDoubleNode(0, n) )
2428     : n;
2429 }
2430 
2431 //=============================================================================
2432 // Generate a fast path/slow path idiom.  Graph looks like:
2433 // [foo] indicates that 'foo' is a parameter
2434 //
2435 //              [in]     NULL
2436 //                 \    /
2437 //                  CmpP
2438 //                  Bool ne
2439 //                   If
2440 //                  /  \
2441 //              True    False-<2>
2442 //              / |
2443 //             /  cast_not_null
2444 //           Load  |    |   ^
2445 //        [fast_test]   |   |
2446 // gvn to   opt_test    |   |
2447 //          /    \      |  <1>
2448 //      True     False  |
2449 //        |         \\  |
2450 //   [slow_call]     \[fast_result]
2451 //    Ctl   Val       \      \
2452 //     |               \      \
2453 //    Catch       <1>   \      \
2454 //   /    \        ^     \      \
2455 //  Ex    No_Ex    |      \      \
2456 //  |       \   \  |       \ <2>  \
2457 //  ...      \  [slow_res] |  |    \   [null_result]
2458 //            \         \--+--+---  |  |
2459 //             \           | /    \ | /
2460 //              --------Region     Phi
2461 //
2462 //=============================================================================
2463 // Code is structured as a series of driver functions all called 'do_XXX' that
2464 // call a set of helper functions.  Helper functions first, then drivers.
2465 
2466 //------------------------------null_check_oop---------------------------------
2467 // Null check oop.  Set null-path control into Region in slot 3.
2468 // Make a cast-not-nullness use the other not-null control.  Return cast.
2469 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2470                                bool never_see_null,
2471                                bool safe_for_replace,
2472                                bool speculative) {
2473   // Initial NULL check taken path
2474   (*null_control) = top();
2475   Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative);
2476 
2477   // Generate uncommon_trap:
2478   if (never_see_null && (*null_control) != top()) {
2479     // If we see an unexpected null at a check-cast we record it and force a
2480     // recompile; the offending check-cast will be compiled to handle NULLs.
2481     // If we see more than one offending BCI, then all checkcasts in the
2482     // method will be compiled to handle NULLs.
2483     PreserveJVMState pjvms(this);
2484     set_control(*null_control);
2485     replace_in_map(value, null());
2486     Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative);
2487     uncommon_trap(reason,
2488                   Deoptimization::Action_make_not_entrant);
2489     (*null_control) = top();    // NULL path is dead
2490   }
2491   if ((*null_control) == top() && safe_for_replace) {
2492     replace_in_map(value, cast);
2493   }
2494 
2495   // Cast away null-ness on the result
2496   return cast;
2497 }
2498 
2499 //------------------------------opt_iff----------------------------------------
2500 // Optimize the fast-check IfNode.  Set the fast-path region slot 2.
2501 // Return slow-path control.
2502 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2503   IfNode *opt_iff = _gvn.transform(iff)->as_If();
2504 
2505   // Fast path taken; set region slot 2
2506   Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) );
2507   region->init_req(2,fast_taken); // Capture fast-control
2508 
2509   // Fast path not-taken, i.e. slow path
2510   Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) );
2511   return slow_taken;
2512 }
2513 
2514 //-----------------------------make_runtime_call-------------------------------
2515 Node* GraphKit::make_runtime_call(int flags,
2516                                   const TypeFunc* call_type, address call_addr,
2517                                   const char* call_name,
2518                                   const TypePtr* adr_type,
2519                                   // The following parms are all optional.
2520                                   // The first NULL ends the list.
2521                                   Node* parm0, Node* parm1,
2522                                   Node* parm2, Node* parm3,
2523                                   Node* parm4, Node* parm5,
2524                                   Node* parm6, Node* parm7) {
2525   // Slow-path call
2526   bool is_leaf = !(flags & RC_NO_LEAF);
2527   bool has_io  = (!is_leaf && !(flags & RC_NO_IO));
2528   if (call_name == NULL) {
2529     assert(!is_leaf, "must supply name for leaf");
2530     call_name = OptoRuntime::stub_name(call_addr);
2531   }
2532   CallNode* call;
2533   if (!is_leaf) {
2534     call = new CallStaticJavaNode(call_type, call_addr, call_name,
2535                                            bci(), adr_type);
2536   } else if (flags & RC_NO_FP) {
2537     call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2538   } else {
2539     call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2540   }
2541 
2542   // The following is similar to set_edges_for_java_call,
2543   // except that the memory effects of the call are restricted to AliasIdxRaw.
2544 
2545   // Slow path call has no side-effects, uses few values
2546   bool wide_in  = !(flags & RC_NARROW_MEM);
2547   bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2548 
2549   Node* prev_mem = NULL;
2550   if (wide_in) {
2551     prev_mem = set_predefined_input_for_runtime_call(call);
2552   } else {
2553     assert(!wide_out, "narrow in => narrow out");
2554     Node* narrow_mem = memory(adr_type);
2555     prev_mem = reset_memory();
2556     map()->set_memory(narrow_mem);
2557     set_predefined_input_for_runtime_call(call);
2558   }
2559 
2560   // Hook each parm in order.  Stop looking at the first NULL.
2561   if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2562   if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2563   if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2564   if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2565   if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2566   if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2567   if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2568   if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2569     /* close each nested if ===> */  } } } } } } } }
2570   assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2571 
2572   if (!is_leaf) {
2573     // Non-leaves can block and take safepoints:
2574     add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2575   }
2576   // Non-leaves can throw exceptions:
2577   if (has_io) {
2578     call->set_req(TypeFunc::I_O, i_o());
2579   }
2580 
2581   if (flags & RC_UNCOMMON) {
2582     // Set the count to a tiny probability.  Cf. Estimate_Block_Frequency.
2583     // (An "if" probability corresponds roughly to an unconditional count.
2584     // Sort of.)
2585     call->set_cnt(PROB_UNLIKELY_MAG(4));
2586   }
2587 
2588   Node* c = _gvn.transform(call);
2589   assert(c == call, "cannot disappear");
2590 
2591   if (wide_out) {
2592     // Slow path call has full side-effects.
2593     set_predefined_output_for_runtime_call(call);
2594   } else {
2595     // Slow path call has few side-effects, and/or sets few values.
2596     set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2597   }
2598 
2599   if (has_io) {
2600     set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2601   }
2602   return call;
2603 
2604 }
2605 
2606 //------------------------------merge_memory-----------------------------------
2607 // Merge memory from one path into the current memory state.
2608 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2609   for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2610     Node* old_slice = mms.force_memory();
2611     Node* new_slice = mms.memory2();
2612     if (old_slice != new_slice) {
2613       PhiNode* phi;
2614       if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2615         if (mms.is_empty()) {
2616           // clone base memory Phi's inputs for this memory slice
2617           assert(old_slice == mms.base_memory(), "sanity");
2618           phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C));
2619           _gvn.set_type(phi, Type::MEMORY);
2620           for (uint i = 1; i < phi->req(); i++) {
2621             phi->init_req(i, old_slice->in(i));
2622           }
2623         } else {
2624           phi = old_slice->as_Phi(); // Phi was generated already
2625         }
2626       } else {
2627         phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2628         _gvn.set_type(phi, Type::MEMORY);
2629       }
2630       phi->set_req(new_path, new_slice);
2631       mms.set_memory(phi);
2632     }
2633   }
2634 }
2635 
2636 //------------------------------make_slow_call_ex------------------------------
2637 // Make the exception handler hookups for the slow call
2638 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2639   if (stopped())  return;
2640 
2641   // Make a catch node with just two handlers:  fall-through and catch-all
2642   Node* i_o  = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2643   Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) );
2644   Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2645   Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index,    CatchProjNode::no_handler_bci) );
2646 
2647   { PreserveJVMState pjvms(this);
2648     set_control(excp);
2649     set_i_o(i_o);
2650 
2651     if (excp != top()) {
2652       if (deoptimize) {
2653         // Deoptimize if an exception is caught. Don't construct exception state in this case.
2654         uncommon_trap(Deoptimization::Reason_unhandled,
2655                       Deoptimization::Action_none);
2656       } else {
2657         // Create an exception state also.
2658         // Use an exact type if the caller has a specific exception.
2659         const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2660         Node*       ex_oop  = new CreateExNode(ex_type, control(), i_o);
2661         add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2662       }
2663     }
2664   }
2665 
2666   // Get the no-exception control from the CatchNode.
2667   set_control(norm);
2668 }
2669 
2670 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) {
2671   Node* cmp = NULL;
2672   switch(bt) {
2673   case T_INT: cmp = new CmpINode(in1, in2); break;
2674   case T_ADDRESS: cmp = new CmpPNode(in1, in2); break;
2675   default: fatal("unexpected comparison type %s", type2name(bt));
2676   }
2677   gvn->transform(cmp);
2678   Node* bol = gvn->transform(new BoolNode(cmp, test));
2679   IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN);
2680   gvn->transform(iff);
2681   if (!bol->is_Con()) gvn->record_for_igvn(iff);
2682   return iff;
2683 }
2684 
2685 
2686 //-------------------------------gen_subtype_check-----------------------------
2687 // Generate a subtyping check.  Takes as input the subtype and supertype.
2688 // Returns 2 values: sets the default control() to the true path and returns
2689 // the false path.  Only reads invariant memory; sets no (visible) memory.
2690 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2691 // but that's not exposed to the optimizer.  This call also doesn't take in an
2692 // Object; if you wish to check an Object you need to load the Object's class
2693 // prior to coming here.
2694 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) {
2695   Compile* C = gvn->C;
2696 
2697   if ((*ctrl)->is_top()) {
2698     return C->top();
2699   }
2700 
2701   // Fast check for identical types, perhaps identical constants.
2702   // The types can even be identical non-constants, in cases
2703   // involving Array.newInstance, Object.clone, etc.
2704   if (subklass == superklass)
2705     return C->top();             // false path is dead; no test needed.
2706 
2707   if (gvn->type(superklass)->singleton()) {
2708     ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass();
2709     ciKlass* subk   = gvn->type(subklass)->is_klassptr()->klass();
2710 
2711     // In the common case of an exact superklass, try to fold up the
2712     // test before generating code.  You may ask, why not just generate
2713     // the code and then let it fold up?  The answer is that the generated
2714     // code will necessarily include null checks, which do not always
2715     // completely fold away.  If they are also needless, then they turn
2716     // into a performance loss.  Example:
2717     //    Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2718     // Here, the type of 'fa' is often exact, so the store check
2719     // of fa[1]=x will fold up, without testing the nullness of x.
2720     switch (C->static_subtype_check(superk, subk)) {
2721     case Compile::SSC_always_false:
2722       {
2723         Node* always_fail = *ctrl;
2724         *ctrl = gvn->C->top();
2725         return always_fail;
2726       }
2727     case Compile::SSC_always_true:
2728       return C->top();
2729     case Compile::SSC_easy_test:
2730       {
2731         // Just do a direct pointer compare and be done.
2732         IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2733         *ctrl = gvn->transform(new IfTrueNode(iff));
2734         return gvn->transform(new IfFalseNode(iff));
2735       }
2736     case Compile::SSC_full_test:
2737       break;
2738     default:
2739       ShouldNotReachHere();
2740     }
2741   }
2742 
2743   // %%% Possible further optimization:  Even if the superklass is not exact,
2744   // if the subklass is the unique subtype of the superklass, the check
2745   // will always succeed.  We could leave a dependency behind to ensure this.
2746 
2747   // First load the super-klass's check-offset
2748   Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2749   Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr()));
2750   Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2751   int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2752   bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2753 
2754   // Load from the sub-klass's super-class display list, or a 1-word cache of
2755   // the secondary superclass list, or a failing value with a sentinel offset
2756   // if the super-klass is an interface or exceptionally deep in the Java
2757   // hierarchy and we have to scan the secondary superclass list the hard way.
2758   // Worst-case type is a little odd: NULL is allowed as a result (usually
2759   // klass loads can never produce a NULL).
2760   Node *chk_off_X = chk_off;
2761 #ifdef _LP64
2762   chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X));
2763 #endif
2764   Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X));
2765   // For some types like interfaces the following loadKlass is from a 1-word
2766   // cache which is mutable so can't use immutable memory.  Other
2767   // types load from the super-class display table which is immutable.
2768   m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr()));
2769   Node *kmem = might_be_cache ? m : C->immutable_memory();
2770   Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::BOTTOM));
2771 
2772   // Compile speed common case: ARE a subtype and we canNOT fail
2773   if( superklass == nkls )
2774     return C->top();             // false path is dead; no test needed.
2775 
2776   // See if we get an immediate positive hit.  Happens roughly 83% of the
2777   // time.  Test to see if the value loaded just previously from the subklass
2778   // is exactly the superklass.
2779   IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
2780   Node *iftrue1 = gvn->transform( new IfTrueNode (iff1));
2781   *ctrl = gvn->transform(new IfFalseNode(iff1));
2782 
2783   // Compile speed common case: Check for being deterministic right now.  If
2784   // chk_off is a constant and not equal to cacheoff then we are NOT a
2785   // subklass.  In this case we need exactly the 1 test above and we can
2786   // return those results immediately.
2787   if (!might_be_cache) {
2788     Node* not_subtype_ctrl = *ctrl;
2789     *ctrl = iftrue1; // We need exactly the 1 test above
2790     return not_subtype_ctrl;
2791   }
2792 
2793   // Gather the various success & failures here
2794   RegionNode *r_ok_subtype = new RegionNode(4);
2795   gvn->record_for_igvn(r_ok_subtype);
2796   RegionNode *r_not_subtype = new RegionNode(3);
2797   gvn->record_for_igvn(r_not_subtype);
2798 
2799   r_ok_subtype->init_req(1, iftrue1);
2800 
2801   // Check for immediate negative hit.  Happens roughly 11% of the time (which
2802   // is roughly 63% of the remaining cases).  Test to see if the loaded
2803   // check-offset points into the subklass display list or the 1-element
2804   // cache.  If it points to the display (and NOT the cache) and the display
2805   // missed then it's not a subtype.
2806   Node *cacheoff = gvn->intcon(cacheoff_con);
2807   IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
2808   r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2)));
2809   *ctrl = gvn->transform(new IfFalseNode(iff2));
2810 
2811   // Check for self.  Very rare to get here, but it is taken 1/3 the time.
2812   // No performance impact (too rare) but allows sharing of secondary arrays
2813   // which has some footprint reduction.
2814   IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
2815   r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3)));
2816   *ctrl = gvn->transform(new IfFalseNode(iff3));
2817 
2818   // -- Roads not taken here: --
2819   // We could also have chosen to perform the self-check at the beginning
2820   // of this code sequence, as the assembler does.  This would not pay off
2821   // the same way, since the optimizer, unlike the assembler, can perform
2822   // static type analysis to fold away many successful self-checks.
2823   // Non-foldable self checks work better here in second position, because
2824   // the initial primary superclass check subsumes a self-check for most
2825   // types.  An exception would be a secondary type like array-of-interface,
2826   // which does not appear in its own primary supertype display.
2827   // Finally, we could have chosen to move the self-check into the
2828   // PartialSubtypeCheckNode, and from there out-of-line in a platform
2829   // dependent manner.  But it is worthwhile to have the check here,
2830   // where it can be perhaps be optimized.  The cost in code space is
2831   // small (register compare, branch).
2832 
2833   // Now do a linear scan of the secondary super-klass array.  Again, no real
2834   // performance impact (too rare) but it's gotta be done.
2835   // Since the code is rarely used, there is no penalty for moving it
2836   // out of line, and it can only improve I-cache density.
2837   // The decision to inline or out-of-line this final check is platform
2838   // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2839   Node* psc = gvn->transform(
2840     new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2841 
2842   IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2843   r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4)));
2844   r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4)));
2845 
2846   // Return false path; set default control to true path.
2847   *ctrl = gvn->transform(r_ok_subtype);
2848   return gvn->transform(r_not_subtype);
2849 }
2850 
2851 // Profile-driven exact type check:
2852 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2853                                     float prob,
2854                                     Node* *casted_receiver) {
2855   const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2856   Node* recv_klass = load_object_klass(receiver);
2857   Node* fail = type_check(recv_klass, tklass, prob);
2858   const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2859   assert(recv_xtype->klass_is_exact(), "");
2860 
2861   // Subsume downstream occurrences of receiver with a cast to
2862   // recv_xtype, since now we know what the type will be.
2863   Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
2864   Node* res = _gvn.transform(cast);
2865   if (recv_xtype->is_valuetypeptr()) {
2866     assert(!gvn().type(res)->is_ptr()->maybe_null(), "should never be null");
2867     res = ValueTypeNode::make_from_oop(this, res, recv_xtype->value_klass());
2868   }
2869 
2870   (*casted_receiver) = res;
2871   // (User must make the replace_in_map call.)
2872 
2873   return fail;
2874 }
2875 
2876 Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass,
2877                            float prob) {
2878   Node* want_klass = makecon(tklass);
2879   Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass));
2880   Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) );
2881   IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2882   set_control(  _gvn.transform( new IfTrueNode (iff)));
2883   Node* fail = _gvn.transform( new IfFalseNode(iff));
2884   return fail;
2885 }
2886 
2887 
2888 //------------------------------seems_never_null-------------------------------
2889 // Use null_seen information if it is available from the profile.
2890 // If we see an unexpected null at a type check we record it and force a
2891 // recompile; the offending check will be recompiled to handle NULLs.
2892 // If we see several offending BCIs, then all checks in the
2893 // method will be recompiled.
2894 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
2895   speculating = !_gvn.type(obj)->speculative_maybe_null();
2896   Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
2897   if (UncommonNullCast               // Cutout for this technique
2898       && obj != null()               // And not the -Xcomp stupid case?
2899       && !too_many_traps(reason)
2900       ) {
2901     if (speculating) {
2902       return true;
2903     }
2904     if (data == NULL)
2905       // Edge case:  no mature data.  Be optimistic here.
2906       return true;
2907     // If the profile has not seen a null, assume it won't happen.
2908     assert(java_bc() == Bytecodes::_checkcast ||
2909            java_bc() == Bytecodes::_instanceof ||
2910            java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2911     return !data->as_BitData()->null_seen();
2912   }
2913   speculating = false;
2914   return false;
2915 }
2916 
2917 //------------------------maybe_cast_profiled_receiver-------------------------
2918 // If the profile has seen exactly one type, narrow to exactly that type.
2919 // Subsequent type checks will always fold up.
2920 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2921                                              ciKlass* require_klass,
2922                                              ciKlass* spec_klass,
2923                                              bool safe_for_replace) {
2924   if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2925 
2926   Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL);
2927 
2928   // Make sure we haven't already deoptimized from this tactic.
2929   if (too_many_traps(reason) || too_many_recompiles(reason))
2930     return NULL;
2931 
2932   // (No, this isn't a call, but it's enough like a virtual call
2933   // to use the same ciMethod accessor to get the profile info...)
2934   // If we have a speculative type use it instead of profiling (which
2935   // may not help us)
2936   ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass;
2937   if (exact_kls != NULL) {// no cast failures here
2938     if (require_klass == NULL ||
2939         C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) {
2940       // If we narrow the type to match what the type profile sees or
2941       // the speculative type, we can then remove the rest of the
2942       // cast.
2943       // This is a win, even if the exact_kls is very specific,
2944       // because downstream operations, such as method calls,
2945       // will often benefit from the sharper type.
2946       Node* exact_obj = not_null_obj; // will get updated in place...
2947       Node* slow_ctl  = type_check_receiver(exact_obj, exact_kls, 1.0,
2948                                             &exact_obj);
2949       { PreserveJVMState pjvms(this);
2950         set_control(slow_ctl);
2951         uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
2952       }
2953       if (safe_for_replace) {
2954         replace_in_map(not_null_obj, exact_obj);
2955       }
2956       return exact_obj;
2957     }
2958     // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us.
2959   }
2960 
2961   return NULL;
2962 }
2963 
2964 /**
2965  * Cast obj to type and emit guard unless we had too many traps here
2966  * already
2967  *
2968  * @param obj       node being casted
2969  * @param type      type to cast the node to
2970  * @param not_null  true if we know node cannot be null
2971  */
2972 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
2973                                         ciKlass* type,
2974                                         bool not_null) {
2975   if (stopped()) {
2976     return obj;
2977   }
2978 
2979   // type == NULL if profiling tells us this object is always null
2980   if (type != NULL) {
2981     Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
2982     Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check;
2983 
2984     if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) &&
2985         !too_many_traps(class_reason) &&
2986         !too_many_recompiles(class_reason)) {
2987       Node* not_null_obj = NULL;
2988       // not_null is true if we know the object is not null and
2989       // there's no need for a null check
2990       if (!not_null) {
2991         Node* null_ctl = top();
2992         not_null_obj = null_check_oop(obj, &null_ctl, true, true, true);
2993         assert(null_ctl->is_top(), "no null control here");
2994       } else {
2995         not_null_obj = obj;
2996       }
2997 
2998       Node* exact_obj = not_null_obj;
2999       ciKlass* exact_kls = type;
3000       Node* slow_ctl  = type_check_receiver(exact_obj, exact_kls, 1.0,
3001                                             &exact_obj);
3002       {
3003         PreserveJVMState pjvms(this);
3004         set_control(slow_ctl);
3005         uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
3006       }
3007       replace_in_map(not_null_obj, exact_obj);
3008       obj = exact_obj;
3009     }
3010   } else {
3011     if (!too_many_traps(Deoptimization::Reason_null_assert) &&
3012         !too_many_recompiles(Deoptimization::Reason_null_assert)) {
3013       Node* exact_obj = null_assert(obj);
3014       replace_in_map(obj, exact_obj);
3015       obj = exact_obj;
3016     }
3017   }
3018   return obj;
3019 }
3020 
3021 //-------------------------------gen_instanceof--------------------------------
3022 // Generate an instance-of idiom.  Used by both the instance-of bytecode
3023 // and the reflective instance-of call.
3024 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
3025   kill_dead_locals();           // Benefit all the uncommon traps
3026   assert( !stopped(), "dead parse path should be checked in callers" );
3027   assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
3028          "must check for not-null not-dead klass in callers");
3029 
3030   // Make the merge point
3031   enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
3032   RegionNode* region = new RegionNode(PATH_LIMIT);
3033   Node*       phi    = new PhiNode(region, TypeInt::BOOL);
3034   C->set_has_split_ifs(true); // Has chance for split-if optimization
3035 
3036   ciProfileData* data = NULL;
3037   if (java_bc() == Bytecodes::_instanceof) {  // Only for the bytecode
3038     data = method()->method_data()->bci_to_data(bci());
3039   }
3040   bool speculative_not_null = false;
3041   bool never_see_null = (ProfileDynamicTypes  // aggressive use of profile
3042                          && seems_never_null(obj, data, speculative_not_null));
3043   bool is_value = obj->is_ValueType();
3044 
3045   // Null check; get casted pointer; set region slot 3
3046   Node* null_ctl = top();
3047   Node* not_null_obj = is_value ? obj : null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3048 
3049   // If not_null_obj is dead, only null-path is taken
3050   if (stopped()) {              // Doing instance-of on a NULL?
3051     set_control(null_ctl);
3052     return intcon(0);
3053   }
3054   region->init_req(_null_path, null_ctl);
3055   phi   ->init_req(_null_path, intcon(0)); // Set null path value
3056   if (null_ctl == top()) {
3057     // Do this eagerly, so that pattern matches like is_diamond_phi
3058     // will work even during parsing.
3059     assert(_null_path == PATH_LIMIT-1, "delete last");
3060     region->del_req(_null_path);
3061     phi   ->del_req(_null_path);
3062   }
3063 
3064   // Do we know the type check always succeed?
3065   if (!is_value) {
3066     bool known_statically = false;
3067     if (_gvn.type(superklass)->singleton()) {
3068       ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
3069       ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
3070       if (subk != NULL && subk->is_loaded()) {
3071         int static_res = C->static_subtype_check(superk, subk);
3072         known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3073       }
3074     }
3075 
3076     if (!known_statically) {
3077       const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3078       // We may not have profiling here or it may not help us. If we
3079       // have a speculative type use it to perform an exact cast.
3080       ciKlass* spec_obj_type = obj_type->speculative_type();
3081       if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
3082         Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
3083         if (stopped()) {            // Profile disagrees with this path.
3084           set_control(null_ctl);    // Null is the only remaining possibility.
3085           return intcon(0);
3086         }
3087         if (cast_obj != NULL &&
3088             // A value that's sometimes null is not something we can optimize well
3089             !(cast_obj->is_ValueType() && null_ctl != top())) {
3090           not_null_obj = cast_obj;
3091           is_value = not_null_obj->is_ValueType();
3092         }
3093       }
3094     }
3095   }
3096 
3097   // Load the object's klass
3098   Node* obj_klass = NULL;
3099   if (is_value) {
3100     obj_klass = makecon(TypeKlassPtr::make(_gvn.type(not_null_obj)->is_valuetype()->value_klass()));
3101   } else {
3102     obj_klass = load_object_klass(not_null_obj);
3103   }
3104 
3105   // Generate the subtype check
3106   Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
3107 
3108   // Plug in the success path to the general merge in slot 1.
3109   region->init_req(_obj_path, control());
3110   phi   ->init_req(_obj_path, intcon(1));
3111 
3112   // Plug in the failing path to the general merge in slot 2.
3113   region->init_req(_fail_path, not_subtype_ctrl);
3114   phi   ->init_req(_fail_path, intcon(0));
3115 
3116   // Return final merged results
3117   set_control( _gvn.transform(region) );
3118   record_for_igvn(region);
3119 
3120   // If we know the type check always succeeds then we don't use the
3121   // profiling data at this bytecode. Don't lose it, feed it to the
3122   // type system as a speculative type.
3123   if (safe_for_replace && !is_value) {
3124     Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3125     replace_in_map(obj, casted_obj);
3126   }
3127 
3128   return _gvn.transform(phi);
3129 }
3130 
3131 //-------------------------------gen_checkcast---------------------------------
3132 // Generate a checkcast idiom.  Used by both the checkcast bytecode and the
3133 // array store bytecode.  Stack must be as-if BEFORE doing the bytecode so the
3134 // uncommon-trap paths work.  Adjust stack after this call.
3135 // If failure_control is supplied and not null, it is filled in with
3136 // the control edge for the cast failure.  Otherwise, an appropriate
3137 // uncommon trap or exception is thrown.
3138 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
3139                               Node* *failure_control) {
3140   kill_dead_locals();           // Benefit all the uncommon traps
3141   const TypeKlassPtr* tk = _gvn.type(superklass)->is_klassptr();
3142   const TypeOopPtr* toop = TypeOopPtr::make_from_klass(tk->klass());
3143 
3144   bool is_value = obj->is_ValueType();
3145 
3146   // Fast cutout:  Check the case that the cast is vacuously true.
3147   // This detects the common cases where the test will short-circuit
3148   // away completely.  We do this before we perform the null check,
3149   // because if the test is going to turn into zero code, we don't
3150   // want a residual null check left around.  (Causes a slowdown,
3151   // for example, in some objArray manipulations, such as a[i]=a[j].)
3152   if (tk->singleton()) {
3153     ciKlass* klass = NULL;
3154     if (is_value) {
3155       klass = _gvn.type(obj)->is_valuetype()->value_klass();
3156     } else {
3157       const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3158       if (objtp != NULL) {
3159         klass = objtp->klass();
3160       }
3161     }
3162     if (klass != NULL) {
3163       switch (C->static_subtype_check(tk->klass(), klass)) {
3164       case Compile::SSC_always_true:
3165         // If we know the type check always succeed then we don't use
3166         // the profiling data at this bytecode. Don't lose it, feed it
3167         // to the type system as a speculative type.
3168         if (!is_value) {
3169           obj = record_profiled_receiver_for_speculation(obj);
3170           if (toop->is_valuetypeptr()) {
3171             obj = ValueTypeNode::make_from_oop(this, obj, toop->value_klass(), /* buffer_check */ false, /* null2default */ false);
3172           }
3173         }
3174         return obj;
3175       case Compile::SSC_always_false:
3176         // It needs a null check because a null will *pass* the cast check.
3177         // A non-null value will always produce an exception.
3178         if (is_value) {
3179           builtin_throw(Deoptimization::Reason_class_check, makecon(TypeKlassPtr::make(klass)));
3180           return top();
3181         } else {
3182           return null_assert(obj);
3183         }
3184       }
3185     }
3186   }
3187 
3188   ciProfileData* data = NULL;
3189   bool safe_for_replace = false;
3190   if (failure_control == NULL) {        // use MDO in regular case only
3191     assert(java_bc() == Bytecodes::_aastore ||
3192            java_bc() == Bytecodes::_checkcast,
3193            "interpreter profiles type checks only for these BCs");
3194     data = method()->method_data()->bci_to_data(bci());
3195     safe_for_replace = true;
3196   }
3197 
3198   // Make the merge point
3199   enum { _obj_path = 1, _null_path, PATH_LIMIT };
3200   RegionNode* region = new RegionNode(PATH_LIMIT);
3201   Node*       phi    = new PhiNode(region, toop);
3202   C->set_has_split_ifs(true); // Has chance for split-if optimization
3203 
3204   // Use null-cast information if it is available
3205   bool speculative_not_null = false;
3206   bool never_see_null = ((failure_control == NULL)  // regular case only
3207                          && seems_never_null(obj, data, speculative_not_null));
3208 
3209   // Null check; get casted pointer; set region slot 3
3210   Node* null_ctl = top();
3211   Node* not_null_obj = is_value ? obj : null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3212 
3213   // If not_null_obj is dead, only null-path is taken
3214   if (stopped()) {              // Doing instance-of on a NULL?
3215     set_control(null_ctl);
3216     return null();
3217   }
3218   region->init_req(_null_path, null_ctl);
3219   phi   ->init_req(_null_path, null());  // Set null path value
3220   if (null_ctl == top()) {
3221     // Do this eagerly, so that pattern matches like is_diamond_phi
3222     // will work even during parsing.
3223     assert(_null_path == PATH_LIMIT-1, "delete last");
3224     region->del_req(_null_path);
3225     phi   ->del_req(_null_path);
3226   }
3227 
3228   Node* cast_obj = NULL;
3229   if (!is_value && tk->klass_is_exact()) {
3230     // The following optimization tries to statically cast the speculative type of the object
3231     // (for example obtained during profiling) to the type of the superklass and then do a
3232     // dynamic check that the type of the object is what we expect. To work correctly
3233     // for checkcast and aastore the type of superklass should be exact.
3234     const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3235     // We may not have profiling here or it may not help us. If we have
3236     // a speculative type use it to perform an exact cast.
3237     ciKlass* spec_obj_type = obj_type->speculative_type();
3238     if (spec_obj_type != NULL || data != NULL) {
3239       cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3240       if (cast_obj != NULL && cast_obj->is_ValueType()) {
3241         if (null_ctl != top()) {
3242           cast_obj = NULL; // A value that's sometimes null is not something we can optimize well
3243         } else {
3244           return cast_obj;
3245         }
3246       }
3247       if (cast_obj != NULL) {
3248         if (failure_control != NULL) // failure is now impossible
3249           (*failure_control) = top();
3250         // adjust the type of the phi to the exact klass:
3251         phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3252       }
3253     }
3254   }
3255 
3256   if (cast_obj == NULL) {
3257     // Load the object's klass
3258     Node* obj_klass = NULL;
3259     if (is_value) {
3260       obj_klass = makecon(TypeKlassPtr::make(_gvn.type(not_null_obj)->is_valuetype()->value_klass()));
3261     } else {
3262       obj_klass = load_object_klass(not_null_obj);
3263     }
3264 
3265     // Generate the subtype check
3266     Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
3267 
3268     if (is_value) {
3269       not_null_obj = ValueTypePtrNode::make_from_value_type(this, not_null_obj->as_ValueType(), true);
3270     }
3271 
3272     // Plug in success path into the merge
3273     cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3274     // Failure path ends in uncommon trap (or may be dead - failure impossible)
3275     if (failure_control == NULL) {
3276       if (not_subtype_ctrl != top()) { // If failure is possible
3277         PreserveJVMState pjvms(this);
3278         set_control(not_subtype_ctrl);
3279         builtin_throw(Deoptimization::Reason_class_check, obj_klass);
3280       }
3281     } else {
3282       (*failure_control) = not_subtype_ctrl;
3283     }
3284   }
3285 
3286   region->init_req(_obj_path, control());
3287   phi   ->init_req(_obj_path, cast_obj);
3288 
3289   // A merge of NULL or Casted-NotNull obj
3290   Node* res = _gvn.transform(phi);
3291 
3292   // Note I do NOT always 'replace_in_map(obj,result)' here.
3293   //  if( tk->klass()->can_be_primary_super()  )
3294     // This means that if I successfully store an Object into an array-of-String
3295     // I 'forget' that the Object is really now known to be a String.  I have to
3296     // do this because we don't have true union types for interfaces - if I store
3297     // a Baz into an array-of-Interface and then tell the optimizer it's an
3298     // Interface, I forget that it's also a Baz and cannot do Baz-like field
3299     // references to it.  FIX THIS WHEN UNION TYPES APPEAR!
3300   //  replace_in_map( obj, res );
3301 
3302   // Return final merged results
3303   set_control( _gvn.transform(region) );
3304   record_for_igvn(region);
3305 
3306   if (!is_value) {
3307     res = record_profiled_receiver_for_speculation(res);
3308     if (toop->is_valuetypeptr()) {
3309       res = ValueTypeNode::make_from_oop(this, res, toop->value_klass(), /* buffer_check */ false, /* null2default */ false);
3310     }
3311   }
3312   return res;
3313 }
3314 
3315 Node* GraphKit::is_always_locked(Node* obj) {
3316   Node* mark_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3317   Node* mark = make_load(NULL, mark_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
3318   Node* value_mask = _gvn.MakeConX(markOopDesc::always_locked_pattern);
3319   return _gvn.transform(new AndXNode(mark, value_mask));
3320 }
3321 
3322 Node* GraphKit::gen_value_type_test(Node* kls) {
3323   Node* flags_addr = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3324   Node* flags = make_load(NULL, flags_addr, TypeInt::INT, T_INT, MemNode::unordered);
3325   Node* is_value = _gvn.transform(new AndINode(flags, intcon(JVM_ACC_VALUE)));
3326   Node* cmp = _gvn.transform(new CmpINode(is_value, intcon(0)));
3327   return cmp;
3328 }
3329 
3330 // Deoptimize if 'obj' is a value type
3331 void GraphKit::gen_value_type_guard(Node* obj, int nargs) {
3332   assert(EnableValhalla, "should only be used if value types are enabled");
3333   Node* bol = NULL;
3334   if (obj->is_ValueTypeBase()) {
3335     bol = intcon(0);
3336   } else {
3337     Node* is_value = is_always_locked(obj);
3338     Node* value_mask = _gvn.MakeConX(markOopDesc::always_locked_pattern);
3339     Node* cmp = _gvn.transform(new CmpXNode(is_value, value_mask));
3340     bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3341   }
3342   { BuildCutout unless(this, bol, PROB_MAX);
3343     inc_sp(nargs);
3344     uncommon_trap(Deoptimization::Reason_class_check,
3345                   Deoptimization::Action_none);
3346   }
3347 }
3348 
3349 // Deoptimize if 'ary' is flattened or if 'obj' is null and 'ary' is a value type array
3350 void GraphKit::gen_value_type_array_guard(Node* ary, Node* obj, int nargs) {
3351   assert(EnableValhalla, "should only be used if value types are enabled");
3352   // Load array element klass
3353   Node* kls = load_object_klass(ary);
3354   Node* k_adr = basic_plus_adr(kls, in_bytes(ArrayKlass::element_klass_offset()));
3355   Node* elem_klass = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS));
3356   // Check if element is a value type
3357   Node* flags_addr = basic_plus_adr(elem_klass, in_bytes(Klass::access_flags_offset()));
3358   Node* flags = make_load(NULL, flags_addr, TypeInt::INT, T_INT, MemNode::unordered);
3359   Node* is_value_elem = _gvn.transform(new AndINode(flags, intcon(JVM_ACC_VALUE)));
3360 
3361   const Type* objtype = _gvn.type(obj);
3362   if (objtype == TypePtr::NULL_PTR) {
3363     // Object is always null, check if array is a value type array
3364     Node* cmp = _gvn.transform(new CmpINode(is_value_elem, intcon(0)));
3365     Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3366     { BuildCutout unless(this, bol, PROB_MAX);
3367       // TODO just deoptimize for now if we store null to a value type array
3368       inc_sp(nargs);
3369       uncommon_trap(Deoptimization::Reason_array_check,
3370                     Deoptimization::Action_none);
3371     }
3372   } else {
3373     // Check if (is_value_elem && obj_is_null) <=> (!is_value_elem | !obj_is_null == 0)
3374     // TODO what if we later figure out that obj is never null?
3375     Node* not_value = _gvn.transform(new XorINode(is_value_elem, intcon(JVM_ACC_VALUE)));
3376     not_value = _gvn.transform(new ConvI2LNode(not_value));
3377     Node* not_null = _gvn.transform(new CastP2XNode(NULL, obj));
3378     Node* both = _gvn.transform(new OrLNode(not_null, not_value));
3379     Node* cmp  = _gvn.transform(new CmpLNode(both, longcon(0)));
3380     Node* bol  = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3381     { BuildCutout unless(this, bol, PROB_MAX);
3382       // TODO just deoptimize for now if we store null to a value type array
3383       inc_sp(nargs);
3384       uncommon_trap(Deoptimization::Reason_array_check,
3385                     Deoptimization::Action_none);
3386     }
3387   }
3388 }
3389 
3390 Node* GraphKit::load_lh_array_tag(Node* kls) {
3391   Node* lhp = basic_plus_adr(kls, in_bytes(Klass::layout_helper_offset()));
3392   Node* layout_val = make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3393   return _gvn.transform(new RShiftINode(layout_val, intcon(Klass::_lh_array_tag_shift)));
3394 }
3395 
3396 
3397 Node* GraphKit::gen_lh_array_test(Node* kls, unsigned int lh_value) {
3398   Node* layout_val = load_lh_array_tag(kls);
3399   Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(lh_value)));
3400   return cmp;
3401 }
3402 
3403 
3404 //------------------------------next_monitor-----------------------------------
3405 // What number should be given to the next monitor?
3406 int GraphKit::next_monitor() {
3407   int current = jvms()->monitor_depth()* C->sync_stack_slots();
3408   int next = current + C->sync_stack_slots();
3409   // Keep the toplevel high water mark current:
3410   if (C->fixed_slots() < next)  C->set_fixed_slots(next);
3411   return current;
3412 }
3413 
3414 //------------------------------insert_mem_bar---------------------------------
3415 // Memory barrier to avoid floating things around
3416 // The membar serves as a pinch point between both control and all memory slices.
3417 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3418   MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3419   mb->init_req(TypeFunc::Control, control());
3420   mb->init_req(TypeFunc::Memory,  reset_memory());
3421   Node* membar = _gvn.transform(mb);
3422   set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3423   set_all_memory_call(membar);
3424   return membar;
3425 }
3426 
3427 //-------------------------insert_mem_bar_volatile----------------------------
3428 // Memory barrier to avoid floating things around
3429 // The membar serves as a pinch point between both control and memory(alias_idx).
3430 // If you want to make a pinch point on all memory slices, do not use this
3431 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3432 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3433   // When Parse::do_put_xxx updates a volatile field, it appends a series
3434   // of MemBarVolatile nodes, one for *each* volatile field alias category.
3435   // The first membar is on the same memory slice as the field store opcode.
3436   // This forces the membar to follow the store.  (Bug 6500685 broke this.)
3437   // All the other membars (for other volatile slices, including AliasIdxBot,
3438   // which stands for all unknown volatile slices) are control-dependent
3439   // on the first membar.  This prevents later volatile loads or stores
3440   // from sliding up past the just-emitted store.
3441 
3442   MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3443   mb->set_req(TypeFunc::Control,control());
3444   if (alias_idx == Compile::AliasIdxBot) {
3445     mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3446   } else {
3447     assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3448     mb->set_req(TypeFunc::Memory, memory(alias_idx));
3449   }
3450   Node* membar = _gvn.transform(mb);
3451   set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3452   if (alias_idx == Compile::AliasIdxBot) {
3453     merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3454   } else {
3455     set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3456   }
3457   return membar;
3458 }
3459 
3460 //------------------------------shared_lock------------------------------------
3461 // Emit locking code.
3462 FastLockNode* GraphKit::shared_lock(Node* obj) {
3463   // bci is either a monitorenter bc or InvocationEntryBci
3464   // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3465   assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3466 
3467   if( !GenerateSynchronizationCode )
3468     return NULL;                // Not locking things?
3469 
3470   // We cannot lock on a value type
3471   const TypeOopPtr* objptr = _gvn.type(obj)->make_oopptr();
3472   if (objptr->can_be_value_type()) {
3473     gen_value_type_guard(obj, 1);
3474   }
3475 
3476   if (stopped())                // Dead monitor?
3477     return NULL;
3478 
3479   assert(dead_locals_are_killed(), "should kill locals before sync. point");
3480 
3481   // Box the stack location
3482   Node* box = _gvn.transform(new BoxLockNode(next_monitor()));
3483   Node* mem = reset_memory();
3484 
3485   FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock();
3486   if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) {
3487     // Create the counters for this fast lock.
3488     flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3489   }
3490 
3491   // Create the rtm counters for this fast lock if needed.
3492   flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3493 
3494   // Add monitor to debug info for the slow path.  If we block inside the
3495   // slow path and de-opt, we need the monitor hanging around
3496   map()->push_monitor( flock );
3497 
3498   const TypeFunc *tf = LockNode::lock_type();
3499   LockNode *lock = new LockNode(C, tf);
3500 
3501   lock->init_req( TypeFunc::Control, control() );
3502   lock->init_req( TypeFunc::Memory , mem );
3503   lock->init_req( TypeFunc::I_O    , top() )     ;   // does no i/o
3504   lock->init_req( TypeFunc::FramePtr, frameptr() );
3505   lock->init_req( TypeFunc::ReturnAdr, top() );
3506 
3507   lock->init_req(TypeFunc::Parms + 0, obj);
3508   lock->init_req(TypeFunc::Parms + 1, box);
3509   lock->init_req(TypeFunc::Parms + 2, flock);
3510   add_safepoint_edges(lock);
3511 
3512   lock = _gvn.transform( lock )->as_Lock();
3513 
3514   // lock has no side-effects, sets few values
3515   set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3516 
3517   insert_mem_bar(Op_MemBarAcquireLock);
3518 
3519   // Add this to the worklist so that the lock can be eliminated
3520   record_for_igvn(lock);
3521 
3522 #ifndef PRODUCT
3523   if (PrintLockStatistics) {
3524     // Update the counter for this lock.  Don't bother using an atomic
3525     // operation since we don't require absolute accuracy.
3526     lock->create_lock_counter(map()->jvms());
3527     increment_counter(lock->counter()->addr());
3528   }
3529 #endif
3530 
3531   return flock;
3532 }
3533 
3534 
3535 //------------------------------shared_unlock----------------------------------
3536 // Emit unlocking code.
3537 void GraphKit::shared_unlock(Node* box, Node* obj) {
3538   // bci is either a monitorenter bc or InvocationEntryBci
3539   // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3540   assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3541 
3542   if( !GenerateSynchronizationCode )
3543     return;
3544   if (stopped()) {               // Dead monitor?
3545     map()->pop_monitor();        // Kill monitor from debug info
3546     return;
3547   }
3548   assert(!obj->is_ValueTypeBase(), "should not unlock on value type");
3549 
3550   // Memory barrier to avoid floating things down past the locked region
3551   insert_mem_bar(Op_MemBarReleaseLock);
3552 
3553   const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3554   UnlockNode *unlock = new UnlockNode(C, tf);
3555 #ifdef ASSERT
3556   unlock->set_dbg_jvms(sync_jvms());
3557 #endif
3558   uint raw_idx = Compile::AliasIdxRaw;
3559   unlock->init_req( TypeFunc::Control, control() );
3560   unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3561   unlock->init_req( TypeFunc::I_O    , top() )     ;   // does no i/o
3562   unlock->init_req( TypeFunc::FramePtr, frameptr() );
3563   unlock->init_req( TypeFunc::ReturnAdr, top() );
3564 
3565   unlock->init_req(TypeFunc::Parms + 0, obj);
3566   unlock->init_req(TypeFunc::Parms + 1, box);
3567   unlock = _gvn.transform(unlock)->as_Unlock();
3568 
3569   Node* mem = reset_memory();
3570 
3571   // unlock has no side-effects, sets few values
3572   set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3573 
3574   // Kill monitor from debug info
3575   map()->pop_monitor( );
3576 }
3577 
3578 //-------------------------------get_layout_helper-----------------------------
3579 // If the given klass is a constant or known to be an array,
3580 // fetch the constant layout helper value into constant_value
3581 // and return (Node*)NULL.  Otherwise, load the non-constant
3582 // layout helper value, and return the node which represents it.
3583 // This two-faced routine is useful because allocation sites
3584 // almost always feature constant types.
3585 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3586   const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3587   if (!StressReflectiveCode && inst_klass != NULL) {
3588     ciKlass* klass = inst_klass->klass();
3589     assert(klass != NULL, "klass should not be NULL");
3590     bool    xklass = inst_klass->klass_is_exact();
3591     bool can_be_value_array = false;
3592     if (klass->is_array_klass() && EnableValhalla && ValueArrayFlatten) {
3593       ciKlass* elem = klass->as_array_klass()->element_klass();
3594       can_be_value_array = elem != NULL && (elem->is_java_lang_Object() || elem->is_interface());
3595     }
3596     if (xklass || (klass->is_array_klass() && !can_be_value_array)) {
3597       jint lhelper = klass->layout_helper();
3598       if (lhelper != Klass::_lh_neutral_value) {
3599         constant_value = lhelper;
3600         return (Node*) NULL;
3601       }
3602     }
3603   }
3604   constant_value = Klass::_lh_neutral_value;  // put in a known value
3605   Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3606   return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3607 }
3608 
3609 // We just put in an allocate/initialize with a big raw-memory effect.
3610 // Hook selected additional alias categories on the initialization.
3611 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3612                                 MergeMemNode* init_in_merge,
3613                                 Node* init_out_raw) {
3614   DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3615   assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3616 
3617   Node* prevmem = kit.memory(alias_idx);
3618   init_in_merge->set_memory_at(alias_idx, prevmem);
3619   kit.set_memory(init_out_raw, alias_idx);
3620 }
3621 
3622 //---------------------------set_output_for_allocation-------------------------
3623 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3624                                           const TypeOopPtr* oop_type,
3625                                           bool deoptimize_on_exception) {
3626   int rawidx = Compile::AliasIdxRaw;
3627   alloc->set_req( TypeFunc::FramePtr, frameptr() );
3628   add_safepoint_edges(alloc);
3629   Node* allocx = _gvn.transform(alloc);
3630   set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3631   // create memory projection for i_o
3632   set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3633   make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3634 
3635   // create a memory projection as for the normal control path
3636   Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3637   set_memory(malloc, rawidx);
3638 
3639   // a normal slow-call doesn't change i_o, but an allocation does
3640   // we create a separate i_o projection for the normal control path
3641   set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3642   Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3643 
3644   // put in an initialization barrier
3645   InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3646                                                  rawoop)->as_Initialize();
3647   assert(alloc->initialization() == init,  "2-way macro link must work");
3648   assert(init ->allocation()     == alloc, "2-way macro link must work");
3649   {
3650     // Extract memory strands which may participate in the new object's
3651     // initialization, and source them from the new InitializeNode.
3652     // This will allow us to observe initializations when they occur,
3653     // and link them properly (as a group) to the InitializeNode.
3654     assert(init->in(InitializeNode::Memory) == malloc, "");
3655     MergeMemNode* minit_in = MergeMemNode::make(malloc);
3656     init->set_req(InitializeNode::Memory, minit_in);
3657     record_for_igvn(minit_in); // fold it up later, if possible
3658     Node* minit_out = memory(rawidx);
3659     assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3660     if (oop_type->isa_aryptr()) {
3661       const TypeAryPtr* arytype = oop_type->is_aryptr();
3662       if (arytype->klass()->is_value_array_klass()) {
3663         ciValueArrayKlass* vak = arytype->klass()->as_value_array_klass();
3664         ciValueKlass* vk = vak->element_klass()->as_value_klass();
3665         for (int i = 0, len = vk->nof_nonstatic_fields(); i < len; i++) {
3666           ciField* field = vk->nonstatic_field_at(i);
3667           if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3668             continue;  // do not bother to track really large numbers of fields
3669           int off_in_vt = field->offset() - vk->first_field_offset();
3670           const TypePtr* adr_type = arytype->with_field_offset(off_in_vt)->add_offset(Type::OffsetBot);
3671           int fieldidx = C->get_alias_index(adr_type);
3672           hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3673         }
3674       } else {
3675         const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3676         int            elemidx  = C->get_alias_index(telemref);
3677         hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3678       }
3679     } else if (oop_type->isa_instptr()) {
3680       ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3681       for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3682         ciField* field = ik->nonstatic_field_at(i);
3683         if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3684           continue;  // do not bother to track really large numbers of fields
3685         // Find (or create) the alias category for this field:
3686         int fieldidx = C->alias_type(field)->index();
3687         hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3688       }
3689     }
3690   }
3691 
3692   // Cast raw oop to the real thing...
3693   Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3694   javaoop = _gvn.transform(javaoop);
3695   C->set_recent_alloc(control(), javaoop);
3696   assert(just_allocated_object(control()) == javaoop, "just allocated");
3697 
3698 #ifdef ASSERT
3699   { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3700     assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3701            "Ideal_allocation works");
3702     assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3703            "Ideal_allocation works");
3704     if (alloc->is_AllocateArray()) {
3705       assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3706              "Ideal_allocation works");
3707       assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3708              "Ideal_allocation works");
3709     } else {
3710       assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3711     }
3712   }
3713 #endif //ASSERT
3714 
3715   return javaoop;
3716 }
3717 
3718 //---------------------------new_instance--------------------------------------
3719 // This routine takes a klass_node which may be constant (for a static type)
3720 // or may be non-constant (for reflective code).  It will work equally well
3721 // for either, and the graph will fold nicely if the optimizer later reduces
3722 // the type to a constant.
3723 // The optional arguments are for specialized use by intrinsics:
3724 //  - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3725 //  - If 'return_size_val', report the the total object size to the caller.
3726 //  - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3727 Node* GraphKit::new_instance(Node* klass_node,
3728                              Node* extra_slow_test,
3729                              Node* *return_size_val,
3730                              bool deoptimize_on_exception,
3731                              ValueTypeBaseNode* value_node) {
3732   // Compute size in doublewords
3733   // The size is always an integral number of doublewords, represented
3734   // as a positive bytewise size stored in the klass's layout_helper.
3735   // The layout_helper also encodes (in a low bit) the need for a slow path.
3736   jint  layout_con = Klass::_lh_neutral_value;
3737   Node* layout_val = get_layout_helper(klass_node, layout_con);
3738   bool  layout_is_con = (layout_val == NULL);
3739 
3740   if (extra_slow_test == NULL)  extra_slow_test = intcon(0);
3741   // Generate the initial go-slow test.  It's either ALWAYS (return a
3742   // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3743   // case) a computed value derived from the layout_helper.
3744   Node* initial_slow_test = NULL;
3745   if (layout_is_con) {
3746     assert(!StressReflectiveCode, "stress mode does not use these paths");
3747     bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3748     initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3749   } else {   // reflective case
3750     // This reflective path is used by Unsafe.allocateInstance.
3751     // (It may be stress-tested by specifying StressReflectiveCode.)
3752     // Basically, we want to get into the VM is there's an illegal argument.
3753     Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3754     initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3755     if (extra_slow_test != intcon(0)) {
3756       initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3757     }
3758     // (Macro-expander will further convert this to a Bool, if necessary.)
3759   }
3760 
3761   // Find the size in bytes.  This is easy; it's the layout_helper.
3762   // The size value must be valid even if the slow path is taken.
3763   Node* size = NULL;
3764   if (layout_is_con) {
3765     size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3766   } else {   // reflective case
3767     // This reflective path is used by clone and Unsafe.allocateInstance.
3768     size = ConvI2X(layout_val);
3769 
3770     // Clear the low bits to extract layout_helper_size_in_bytes:
3771     assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3772     Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3773     size = _gvn.transform( new AndXNode(size, mask) );
3774   }
3775   if (return_size_val != NULL) {
3776     (*return_size_val) = size;
3777   }
3778 
3779   // This is a precise notnull oop of the klass.
3780   // (Actually, it need not be precise if this is a reflective allocation.)
3781   // It's what we cast the result to.
3782   const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3783   if (!tklass)  tklass = TypeKlassPtr::OBJECT;
3784   const TypeOopPtr* oop_type = tklass->as_instance_type();
3785 
3786   // Now generate allocation code
3787 
3788   // The entire memory state is needed for slow path of the allocation
3789   // since GC and deoptimization can happen.
3790   Node *mem = reset_memory();
3791   set_all_memory(mem); // Create new memory state
3792 
3793   AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3794                                          control(), mem, i_o(),
3795                                          size, klass_node,
3796                                          initial_slow_test, value_node);
3797 
3798   return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3799 }
3800 
3801 // With compressed oops, the 64 bit init value for non flattened value
3802 // arrays is built from 2 32 bit compressed oops
3803 static Node* raw_default_for_coops(Node* default_value, GraphKit& kit) {
3804   Node* lower = kit.gvn().transform(new CastP2XNode(kit.control(), default_value));
3805   Node* upper = kit.gvn().transform(new LShiftLNode(lower, kit.intcon(32)));
3806   return kit.gvn().transform(new OrLNode(lower, upper));
3807 }
3808 
3809 //-------------------------------new_array-------------------------------------
3810 // helper for newarray and anewarray
3811 // The 'length' parameter is (obviously) the length of the array.
3812 // See comments on new_instance for the meaning of the other arguments.
3813 Node* GraphKit::new_array(Node* klass_node,     // array klass (maybe variable)
3814                           Node* length,         // number of array elements
3815                           int   nargs,          // number of arguments to push back for uncommon trap
3816                           Node* *return_size_val,
3817                           bool deoptimize_on_exception) {
3818   jint  layout_con = Klass::_lh_neutral_value;
3819   Node* layout_val = get_layout_helper(klass_node, layout_con);
3820   bool  layout_is_con = (layout_val == NULL);
3821 
3822   if (!layout_is_con && !StressReflectiveCode &&
3823       !too_many_traps(Deoptimization::Reason_class_check)) {
3824     // This is a reflective array creation site.
3825     // Optimistically assume that it is a subtype of Object[],
3826     // so that we can fold up all the address arithmetic.
3827     layout_con = Klass::array_layout_helper(T_OBJECT);
3828     Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
3829     Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
3830     { BuildCutout unless(this, bol_lh, PROB_MAX);
3831       inc_sp(nargs);
3832       uncommon_trap(Deoptimization::Reason_class_check,
3833                     Deoptimization::Action_maybe_recompile);
3834     }
3835     layout_val = NULL;
3836     layout_is_con = true;
3837   }
3838 
3839   // Generate the initial go-slow test.  Make sure we do not overflow
3840   // if length is huge (near 2Gig) or negative!  We do not need
3841   // exact double-words here, just a close approximation of needed
3842   // double-words.  We can't add any offset or rounding bits, lest we
3843   // take a size -1 of bytes and make it positive.  Use an unsigned
3844   // compare, so negative sizes look hugely positive.
3845   int fast_size_limit = FastAllocateSizeLimit;
3846   if (layout_is_con) {
3847     assert(!StressReflectiveCode, "stress mode does not use these paths");
3848     // Increase the size limit if we have exact knowledge of array type.
3849     int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3850     fast_size_limit <<= MAX2(LogBytesPerLong - log2_esize, 0);
3851   }
3852 
3853   Node* initial_slow_cmp  = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
3854   Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
3855 
3856   // --- Size Computation ---
3857   // array_size = round_to_heap(array_header + (length << elem_shift));
3858   // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
3859   // and align_to(x, y) == ((x + y-1) & ~(y-1))
3860   // The rounding mask is strength-reduced, if possible.
3861   int round_mask = MinObjAlignmentInBytes - 1;
3862   Node* header_size = NULL;
3863   int   header_size_min  = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3864   // (T_BYTE has the weakest alignment and size restrictions...)
3865   if (layout_is_con) {
3866     int       hsize  = Klass::layout_helper_header_size(layout_con);
3867     int       eshift = Klass::layout_helper_log2_element_size(layout_con);
3868     BasicType etype  = Klass::layout_helper_element_type(layout_con);
3869     bool is_value_array = Klass::layout_helper_is_valueArray(layout_con);
3870     if ((round_mask & ~right_n_bits(eshift)) == 0)
3871       round_mask = 0;  // strength-reduce it if it goes away completely
3872     assert(is_value_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3873     assert(header_size_min <= hsize, "generic minimum is smallest");
3874     header_size_min = hsize;
3875     header_size = intcon(hsize + round_mask);
3876   } else {
3877     Node* hss   = intcon(Klass::_lh_header_size_shift);
3878     Node* hsm   = intcon(Klass::_lh_header_size_mask);
3879     Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) );
3880     hsize       = _gvn.transform( new AndINode(hsize, hsm) );
3881     Node* mask  = intcon(round_mask);
3882     header_size = _gvn.transform( new AddINode(hsize, mask) );
3883   }
3884 
3885   Node* elem_shift = NULL;
3886   if (layout_is_con) {
3887     int eshift = Klass::layout_helper_log2_element_size(layout_con);
3888     if (eshift != 0)
3889       elem_shift = intcon(eshift);
3890   } else {
3891     // There is no need to mask or shift this value.
3892     // The semantics of LShiftINode include an implicit mask to 0x1F.
3893     assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3894     elem_shift = layout_val;
3895   }
3896 
3897   // Transition to native address size for all offset calculations:
3898   Node* lengthx = ConvI2X(length);
3899   Node* headerx = ConvI2X(header_size);
3900 #ifdef _LP64
3901   { const TypeInt* tilen = _gvn.find_int_type(length);
3902     if (tilen != NULL && tilen->_lo < 0) {
3903       // Add a manual constraint to a positive range.  Cf. array_element_address.
3904       jint size_max = fast_size_limit;
3905       if (size_max > tilen->_hi)  size_max = tilen->_hi;
3906       const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
3907 
3908       // Only do a narrow I2L conversion if the range check passed.
3909       IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
3910       _gvn.transform(iff);
3911       RegionNode* region = new RegionNode(3);
3912       _gvn.set_type(region, Type::CONTROL);
3913       lengthx = new PhiNode(region, TypeLong::LONG);
3914       _gvn.set_type(lengthx, TypeLong::LONG);
3915 
3916       // Range check passed. Use ConvI2L node with narrow type.
3917       Node* passed = IfFalse(iff);
3918       region->init_req(1, passed);
3919       // Make I2L conversion control dependent to prevent it from
3920       // floating above the range check during loop optimizations.
3921       lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
3922 
3923       // Range check failed. Use ConvI2L with wide type because length may be invalid.
3924       region->init_req(2, IfTrue(iff));
3925       lengthx->init_req(2, ConvI2X(length));
3926 
3927       set_control(region);
3928       record_for_igvn(region);
3929       record_for_igvn(lengthx);
3930     }
3931   }
3932 #endif
3933 
3934   // Combine header size (plus rounding) and body size.  Then round down.
3935   // This computation cannot overflow, because it is used only in two
3936   // places, one where the length is sharply limited, and the other
3937   // after a successful allocation.
3938   Node* abody = lengthx;
3939   if (elem_shift != NULL)
3940     abody     = _gvn.transform( new LShiftXNode(lengthx, elem_shift) );
3941   Node* size  = _gvn.transform( new AddXNode(headerx, abody) );
3942   if (round_mask != 0) {
3943     Node* mask = MakeConX(~round_mask);
3944     size       = _gvn.transform( new AndXNode(size, mask) );
3945   }
3946   // else if round_mask == 0, the size computation is self-rounding
3947 
3948   if (return_size_val != NULL) {
3949     // This is the size
3950     (*return_size_val) = size;
3951   }
3952 
3953   // Now generate allocation code
3954 
3955   // The entire memory state is needed for slow path of the allocation
3956   // since GC and deoptimization can happen.
3957   Node *mem = reset_memory();
3958   set_all_memory(mem); // Create new memory state
3959 
3960   if (initial_slow_test->is_Bool()) {
3961     // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3962     initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3963   }
3964 
3965   const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3966   const TypeAryPtr* ary_ptr = ary_type->isa_aryptr();
3967   const Type* elem = NULL;
3968   ciKlass* elem_klass = NULL;
3969   if (ary_ptr != NULL) {
3970     elem = ary_ptr->elem();
3971     elem_klass = ary_ptr->klass()->as_array_klass()->element_klass();
3972   }
3973   Node* default_value = NULL;
3974   Node* raw_default_value = NULL;
3975   if (elem != NULL && elem->make_ptr()) {
3976     if (elem_klass != NULL && elem_klass->is_valuetype()) {
3977       ciValueKlass* vk = elem_klass->as_value_klass();
3978       if (!vk->flatten_array()) {
3979         default_value = ValueTypeNode::default_oop(gvn(), vk);
3980         if (elem->isa_narrowoop()) {
3981           default_value = _gvn.transform(new EncodePNode(default_value, elem));
3982           raw_default_value = raw_default_for_coops(default_value, *this);
3983         } else {
3984           raw_default_value = _gvn.transform(new CastP2XNode(control(), default_value));
3985         }
3986       }
3987     }
3988   }
3989 
3990   if (EnableValhalla && (elem == NULL || (elem_klass != NULL && elem_klass->is_java_lang_Object() && !ary_type->klass_is_exact()))) {
3991     assert(raw_default_value == NULL, "shouldn't be set yet");
3992 
3993     // unkown array type, could be a non flattened value array that's
3994     // initialize to a non zero default value
3995 
3996     Node* r = new RegionNode(4);
3997     Node* phi = new PhiNode(r, TypeX_X);
3998 
3999     Node* cmp = gen_lh_array_test(klass_node, Klass::_lh_array_tag_obj_value);
4000     Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
4001     IfNode* iff = create_and_map_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN);
4002     r->init_req(1, _gvn.transform(new IfFalseNode(iff)));
4003     phi->init_req(1, MakeConX(0));
4004     set_control(_gvn.transform(new IfTrueNode(iff)));
4005     Node* k_adr = basic_plus_adr(klass_node, in_bytes(ArrayKlass::element_klass_offset()));
4006     Node* elem_klass = _gvn.transform(LoadKlassNode::make(_gvn, control(), immutable_memory(), k_adr, TypeInstPtr::KLASS));
4007     cmp = gen_value_type_test(elem_klass);
4008     bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
4009     iff = create_and_map_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN);
4010     r->init_req(2, _gvn.transform(new IfTrueNode(iff)));
4011     phi->init_req(2, MakeConX(0));
4012     set_control(_gvn.transform(new IfFalseNode(iff)));
4013 
4014     Node* adr_fixed_block_addr = basic_plus_adr(elem_klass, in_bytes(InstanceKlass::adr_valueklass_fixed_block_offset()));
4015     Node* adr_fixed_block = make_load(control(), adr_fixed_block_addr, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4016 
4017     Node* default_value_offset_addr = basic_plus_adr(adr_fixed_block, in_bytes(ValueKlass::default_value_offset_offset()));
4018     Node* default_value_offset = make_load(control(), default_value_offset_addr, TypeInt::INT, T_INT, MemNode::unordered);
4019 
4020     Node* elem_mirror = load_mirror_from_klass(elem_klass);
4021 
4022     Node* default_value_addr = basic_plus_adr(elem_mirror, ConvI2X(default_value_offset));
4023     const TypePtr* adr_type = _gvn.type(default_value_addr)->is_ptr();
4024     Node* val = access_load_at(elem_mirror, default_value_addr, adr_type, TypeInstPtr::BOTTOM, T_OBJECT, IN_HEAP);
4025 
4026     if (UseCompressedOops) {
4027       val = _gvn.transform(new EncodePNode(val, elem));
4028       val = raw_default_for_coops(val, *this);
4029     } else {
4030       val = _gvn.transform(new CastP2XNode(control(), val));
4031     }
4032     r->init_req(3, control());
4033     phi->init_req(3, val);
4034     set_control(_gvn.transform(r));
4035     raw_default_value = _gvn.transform(phi);
4036   }
4037 
4038   // Create the AllocateArrayNode and its result projections
4039   AllocateArrayNode* alloc
4040     = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
4041                             control(), mem, i_o(),
4042                             size, klass_node,
4043                             initial_slow_test,
4044                             length, default_value,
4045                             raw_default_value);
4046 
4047   // Cast to correct type.  Note that the klass_node may be constant or not,
4048   // and in the latter case the actual array type will be inexact also.
4049   // (This happens via a non-constant argument to inline_native_newArray.)
4050   // In any case, the value of klass_node provides the desired array type.
4051   const TypeInt* length_type = _gvn.find_int_type(length);
4052   if (ary_type->isa_aryptr() && length_type != NULL) {
4053     // Try to get a better type than POS for the size
4054     ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4055   }
4056 
4057   Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
4058 
4059   // Cast length on remaining path to be as narrow as possible
4060   if (map()->find_edge(length) >= 0) {
4061     Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
4062     if (ccast != length) {
4063       _gvn.set_type_bottom(ccast);
4064       record_for_igvn(ccast);
4065       replace_in_map(length, ccast);
4066     }
4067   }
4068 
4069   return javaoop;
4070 }
4071 
4072 // The following "Ideal_foo" functions are placed here because they recognize
4073 // the graph shapes created by the functions immediately above.
4074 
4075 //---------------------------Ideal_allocation----------------------------------
4076 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
4077 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
4078   if (ptr == NULL) {     // reduce dumb test in callers
4079     return NULL;
4080   }
4081   if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
4082     ptr = ptr->in(1);
4083     if (ptr == NULL) return NULL;
4084   }
4085   // Return NULL for allocations with several casts:
4086   //   j.l.reflect.Array.newInstance(jobject, jint)
4087   //   Object.clone()
4088   // to keep more precise type from last cast.
4089   if (ptr->is_Proj()) {
4090     Node* allo = ptr->in(0);
4091     if (allo != NULL && allo->is_Allocate()) {
4092       return allo->as_Allocate();
4093     }
4094   }
4095   // Report failure to match.
4096   return NULL;
4097 }
4098 
4099 // Fancy version which also strips off an offset (and reports it to caller).
4100 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
4101                                              intptr_t& offset) {
4102   Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
4103   if (base == NULL)  return NULL;
4104   return Ideal_allocation(base, phase);
4105 }
4106 
4107 // Trace Initialize <- Proj[Parm] <- Allocate
4108 AllocateNode* InitializeNode::allocation() {
4109   Node* rawoop = in(InitializeNode::RawAddress);
4110   if (rawoop->is_Proj()) {
4111     Node* alloc = rawoop->in(0);
4112     if (alloc->is_Allocate()) {
4113       return alloc->as_Allocate();
4114     }
4115   }
4116   return NULL;
4117 }
4118 
4119 // Trace Allocate -> Proj[Parm] -> Initialize
4120 InitializeNode* AllocateNode::initialization() {
4121   ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress);
4122   if (rawoop == NULL)  return NULL;
4123   for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
4124     Node* init = rawoop->fast_out(i);
4125     if (init->is_Initialize()) {
4126       assert(init->as_Initialize()->allocation() == this, "2-way link");
4127       return init->as_Initialize();
4128     }
4129   }
4130   return NULL;
4131 }
4132 
4133 //----------------------------- loop predicates ---------------------------
4134 
4135 //------------------------------add_predicate_impl----------------------------
4136 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
4137   // Too many traps seen?
4138   if (too_many_traps(reason)) {
4139 #ifdef ASSERT
4140     if (TraceLoopPredicate) {
4141       int tc = C->trap_count(reason);
4142       tty->print("too many traps=%s tcount=%d in ",
4143                     Deoptimization::trap_reason_name(reason), tc);
4144       method()->print(); // which method has too many predicate traps
4145       tty->cr();
4146     }
4147 #endif
4148     // We cannot afford to take more traps here,
4149     // do not generate predicate.
4150     return;
4151   }
4152 
4153   Node *cont    = _gvn.intcon(1);
4154   Node* opq     = _gvn.transform(new Opaque1Node(C, cont));
4155   Node *bol     = _gvn.transform(new Conv2BNode(opq));
4156   IfNode* iff   = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
4157   Node* iffalse = _gvn.transform(new IfFalseNode(iff));
4158   C->add_predicate_opaq(opq);
4159   {
4160     PreserveJVMState pjvms(this);
4161     set_control(iffalse);
4162     inc_sp(nargs);
4163     uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
4164   }
4165   Node* iftrue = _gvn.transform(new IfTrueNode(iff));
4166   set_control(iftrue);
4167 }
4168 
4169 //------------------------------add_predicate---------------------------------
4170 void GraphKit::add_predicate(int nargs) {
4171   if (UseLoopPredicate) {
4172     add_predicate_impl(Deoptimization::Reason_predicate, nargs);
4173   }
4174   if (UseProfiledLoopPredicate) {
4175     add_predicate_impl(Deoptimization::Reason_profile_predicate, nargs);
4176   }
4177   // loop's limit check predicate should be near the loop.
4178   add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
4179 }
4180 
4181 void GraphKit::sync_kit(IdealKit& ideal) {
4182   set_all_memory(ideal.merged_memory());
4183   set_i_o(ideal.i_o());
4184   set_control(ideal.ctrl());
4185 }
4186 
4187 void GraphKit::final_sync(IdealKit& ideal) {
4188   // Final sync IdealKit and graphKit.
4189   sync_kit(ideal);
4190 }
4191 
4192 Node* GraphKit::load_String_length(Node* ctrl, Node* str) {
4193   Node* len = load_array_length(load_String_value(ctrl, str));
4194   Node* coder = load_String_coder(ctrl, str);
4195   // Divide length by 2 if coder is UTF16
4196   return _gvn.transform(new RShiftINode(len, coder));
4197 }
4198 
4199 Node* GraphKit::load_String_value(Node* ctrl, Node* str) {
4200   int value_offset = java_lang_String::value_offset_in_bytes();
4201   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4202                                                      false, NULL, Type::Offset(0));
4203   const TypePtr* value_field_type = string_type->add_offset(value_offset);
4204   const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4205                                                   TypeAry::make(TypeInt::BYTE, TypeInt::POS),
4206                                                   ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0));
4207   Node* p = basic_plus_adr(str, str, value_offset);
4208   Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4209                               IN_HEAP | C2_CONTROL_DEPENDENT_LOAD);
4210   // String.value field is known to be @Stable.
4211   if (UseImplicitStableValues) {
4212     load = cast_array_to_stable(load, value_type);
4213   }
4214   return load;
4215 }
4216 
4217 Node* GraphKit::load_String_coder(Node* ctrl, Node* str) {
4218   if (!CompactStrings) {
4219     return intcon(java_lang_String::CODER_UTF16);
4220   }
4221   int coder_offset = java_lang_String::coder_offset_in_bytes();
4222   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4223                                                      false, NULL, Type::Offset(0));
4224   const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4225   int coder_field_idx = C->get_alias_index(coder_field_type);
4226   return make_load(ctrl, basic_plus_adr(str, str, coder_offset),
4227                    TypeInt::BYTE, T_BYTE, coder_field_idx, MemNode::unordered);
4228 }
4229 
4230 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) {
4231   int value_offset = java_lang_String::value_offset_in_bytes();
4232   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4233                                                      false, NULL, Type::Offset(0));
4234   const TypePtr* value_field_type = string_type->add_offset(value_offset);
4235   access_store_at(ctrl, str,  basic_plus_adr(str, value_offset), value_field_type,
4236                   value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP);
4237 }
4238 
4239 void GraphKit::store_String_coder(Node* ctrl, Node* str, Node* value) {
4240   int coder_offset = java_lang_String::coder_offset_in_bytes();
4241   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4242                                                      false, NULL, Type::Offset(0));
4243   const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4244   int coder_field_idx = C->get_alias_index(coder_field_type);
4245   store_to_memory(ctrl, basic_plus_adr(str, coder_offset),
4246                   value, T_BYTE, coder_field_idx, MemNode::unordered);
4247 }
4248 
4249 // Capture src and dst memory state with a MergeMemNode
4250 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4251   if (src_type == dst_type) {
4252     // Types are equal, we don't need a MergeMemNode
4253     return memory(src_type);
4254   }
4255   MergeMemNode* merge = MergeMemNode::make(map()->memory());
4256   record_for_igvn(merge); // fold it up later, if possible
4257   int src_idx = C->get_alias_index(src_type);
4258   int dst_idx = C->get_alias_index(dst_type);
4259   merge->set_memory_at(src_idx, memory(src_idx));
4260   merge->set_memory_at(dst_idx, memory(dst_idx));
4261   return merge;
4262 }
4263 
4264 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) {
4265   assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported");
4266   assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type");
4267   // If input and output memory types differ, capture both states to preserve
4268   // the dependency between preceding and subsequent loads/stores.
4269   // For example, the following program:
4270   //  StoreB
4271   //  compress_string
4272   //  LoadB
4273   // has this memory graph (use->def):
4274   //  LoadB -> compress_string -> CharMem
4275   //             ... -> StoreB -> ByteMem
4276   // The intrinsic hides the dependency between LoadB and StoreB, causing
4277   // the load to read from memory not containing the result of the StoreB.
4278   // The correct memory graph should look like this:
4279   //  LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem))
4280   Node* mem = capture_memory(src_type, TypeAryPtr::BYTES);
4281   StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count);
4282   Node* res_mem = _gvn.transform(new SCMemProjNode(str));
4283   set_memory(res_mem, TypeAryPtr::BYTES);
4284   return str;
4285 }
4286 
4287 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) {
4288   assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported");
4289   assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type");
4290   // Capture src and dst memory (see comment in 'compress_string').
4291   Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type);
4292   StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count);
4293   set_memory(_gvn.transform(str), dst_type);
4294 }
4295 
4296 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) {
4297   /**
4298    * int i_char = start;
4299    * for (int i_byte = 0; i_byte < count; i_byte++) {
4300    *   dst[i_char++] = (char)(src[i_byte] & 0xff);
4301    * }
4302    */
4303   add_predicate();
4304   RegionNode* head = new RegionNode(3);
4305   head->init_req(1, control());
4306   gvn().set_type(head, Type::CONTROL);
4307   record_for_igvn(head);
4308 
4309   Node* i_byte = new PhiNode(head, TypeInt::INT);
4310   i_byte->init_req(1, intcon(0));
4311   gvn().set_type(i_byte, TypeInt::INT);
4312   record_for_igvn(i_byte);
4313 
4314   Node* i_char = new PhiNode(head, TypeInt::INT);
4315   i_char->init_req(1, start);
4316   gvn().set_type(i_char, TypeInt::INT);
4317   record_for_igvn(i_char);
4318 
4319   Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES);
4320   gvn().set_type(mem, Type::MEMORY);
4321   record_for_igvn(mem);
4322   set_control(head);
4323   set_memory(mem, TypeAryPtr::BYTES);
4324   Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES);
4325   Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE),
4326                              AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
4327                              false, false, true /* mismatched */);
4328 
4329   IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
4330   head->init_req(2, IfTrue(iff));
4331   mem->init_req(2, st);
4332   i_byte->init_req(2, AddI(i_byte, intcon(1)));
4333   i_char->init_req(2, AddI(i_char, intcon(2)));
4334 
4335   set_control(IfFalse(iff));
4336   set_memory(st, TypeAryPtr::BYTES);
4337 }
4338 
4339 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4340   if (!field->is_constant()) {
4341     return NULL; // Field not marked as constant.
4342   }
4343   ciInstance* holder = NULL;
4344   if (!field->is_static()) {
4345     ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4346     if (const_oop != NULL && const_oop->is_instance()) {
4347       holder = const_oop->as_instance();
4348     }
4349   }
4350   const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4351                                                         /*is_unsigned_load=*/false);
4352   if (con_type != NULL) {
4353     Node* con = makecon(con_type);
4354     if (field->layout_type() == T_VALUETYPE) {
4355       // Load value type from constant oop
4356       con = ValueTypeNode::make_from_oop(this, con, field->type()->as_value_klass());
4357     }
4358     return con;
4359   }
4360   return NULL;
4361 }
4362 
4363 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) {
4364   // Reify the property as a CastPP node in Ideal graph to comply with monotonicity
4365   // assumption of CCP analysis.
4366   return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true)));
4367 }
4368 
4369 //---------------------------load_mirror_from_klass----------------------------
4370 // Given a klass oop, load its java mirror (a java.lang.Class oop).
4371 Node* GraphKit::load_mirror_from_klass(Node* klass) {
4372   Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
4373   Node* load = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4374   // mirror = ((OopHandle)mirror)->resolve();
4375   return access_load(load, TypeInstPtr::MIRROR, T_OBJECT, IN_NATIVE);
4376 }