1 /*
   2  * Copyright (c) 2001, 2025, 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 "asm/register.hpp"
  26 #include "ci/ciFlatArrayKlass.hpp"
  27 #include "ci/ciInlineKlass.hpp"
  28 #include "ci/ciObjArray.hpp"
  29 #include "ci/ciUtilities.hpp"
  30 #include "classfile/javaClasses.hpp"
  31 #include "compiler/compileLog.hpp"
  32 #include "gc/shared/barrierSet.hpp"
  33 #include "gc/shared/c2/barrierSetC2.hpp"
  34 #include "interpreter/interpreter.hpp"
  35 #include "memory/resourceArea.hpp"
  36 #include "oops/flatArrayKlass.hpp"
  37 #include "opto/addnode.hpp"
  38 #include "opto/castnode.hpp"
  39 #include "opto/convertnode.hpp"
  40 #include "opto/graphKit.hpp"
  41 #include "opto/idealKit.hpp"
  42 #include "opto/inlinetypenode.hpp"
  43 #include "opto/intrinsicnode.hpp"
  44 #include "opto/locknode.hpp"
  45 #include "opto/machnode.hpp"
  46 #include "opto/multnode.hpp"
  47 #include "opto/narrowptrnode.hpp"
  48 #include "opto/opaquenode.hpp"
  49 #include "opto/parse.hpp"
  50 #include "opto/rootnode.hpp"
  51 #include "opto/runtime.hpp"
  52 #include "opto/subtypenode.hpp"
  53 #include "runtime/arguments.hpp"
  54 #include "runtime/deoptimization.hpp"
  55 #include "runtime/sharedRuntime.hpp"
  56 #include "runtime/stubRoutines.hpp"
  57 #include "utilities/bitMap.inline.hpp"
  58 #include "utilities/growableArray.hpp"
  59 #include "utilities/powerOfTwo.hpp"
  60 
  61 //----------------------------GraphKit-----------------------------------------
  62 // Main utility constructor.
  63 GraphKit::GraphKit(JVMState* jvms, PhaseGVN* gvn)
  64   : Phase(Phase::Parser),
  65     _env(C->env()),
  66     _gvn((gvn != nullptr) ? *gvn : *C->initial_gvn()),
  67     _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
  68 {
  69   assert(gvn == nullptr || !gvn->is_IterGVN() || gvn->is_IterGVN()->delay_transform(), "delay transform should be enabled");
  70   _exceptions = jvms->map()->next_exception();
  71   if (_exceptions != nullptr)  jvms->map()->set_next_exception(nullptr);
  72   set_jvms(jvms);
  73 #ifdef ASSERT
  74   if (_gvn.is_IterGVN() != nullptr) {
  75     assert(_gvn.is_IterGVN()->delay_transform(), "Transformation must be delayed if IterGVN is used");
  76     // Save the initial size of _for_igvn worklist for verification (see ~GraphKit)
  77     _worklist_size = _gvn.C->igvn_worklist()->size();
  78   }
  79 #endif
  80 }
  81 
  82 // Private constructor for parser.
  83 GraphKit::GraphKit()
  84   : Phase(Phase::Parser),
  85     _env(C->env()),
  86     _gvn(*C->initial_gvn()),
  87     _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
  88 {
  89   _exceptions = nullptr;
  90   set_map(nullptr);
  91   DEBUG_ONLY(_sp = -99);
  92   DEBUG_ONLY(set_bci(-99));
  93 }
  94 
  95 
  96 
  97 //---------------------------clean_stack---------------------------------------
  98 // Clear away rubbish from the stack area of the JVM state.
  99 // This destroys any arguments that may be waiting on the stack.
 100 void GraphKit::clean_stack(int from_sp) {
 101   SafePointNode* map      = this->map();
 102   JVMState*      jvms     = this->jvms();
 103   int            stk_size = jvms->stk_size();
 104   int            stkoff   = jvms->stkoff();
 105   Node*          top      = this->top();
 106   for (int i = from_sp; i < stk_size; i++) {
 107     if (map->in(stkoff + i) != top) {
 108       map->set_req(stkoff + i, top);
 109     }
 110   }
 111 }
 112 
 113 
 114 //--------------------------------sync_jvms-----------------------------------
 115 // Make sure our current jvms agrees with our parse state.
 116 JVMState* GraphKit::sync_jvms() const {
 117   JVMState* jvms = this->jvms();
 118   jvms->set_bci(bci());       // Record the new bci in the JVMState
 119   jvms->set_sp(sp());         // Record the new sp in the JVMState
 120   assert(jvms_in_sync(), "jvms is now in sync");
 121   return jvms;
 122 }
 123 
 124 //--------------------------------sync_jvms_for_reexecute---------------------
 125 // Make sure our current jvms agrees with our parse state.  This version
 126 // uses the reexecute_sp for reexecuting bytecodes.
 127 JVMState* GraphKit::sync_jvms_for_reexecute() {
 128   JVMState* jvms = this->jvms();
 129   jvms->set_bci(bci());          // Record the new bci in the JVMState
 130   jvms->set_sp(reexecute_sp());  // Record the new sp in the JVMState
 131   return jvms;
 132 }
 133 
 134 #ifdef ASSERT
 135 bool GraphKit::jvms_in_sync() const {
 136   Parse* parse = is_Parse();
 137   if (parse == nullptr) {
 138     if (bci() !=      jvms()->bci())          return false;
 139     if (sp()  != (int)jvms()->sp())           return false;
 140     return true;
 141   }
 142   if (jvms()->method() != parse->method())    return false;
 143   if (jvms()->bci()    != parse->bci())       return false;
 144   int jvms_sp = jvms()->sp();
 145   if (jvms_sp          != parse->sp())        return false;
 146   int jvms_depth = jvms()->depth();
 147   if (jvms_depth       != parse->depth())     return false;
 148   return true;
 149 }
 150 
 151 // Local helper checks for special internal merge points
 152 // used to accumulate and merge exception states.
 153 // They are marked by the region's in(0) edge being the map itself.
 154 // Such merge points must never "escape" into the parser at large,
 155 // until they have been handed to gvn.transform.
 156 static bool is_hidden_merge(Node* reg) {
 157   if (reg == nullptr)  return false;
 158   if (reg->is_Phi()) {
 159     reg = reg->in(0);
 160     if (reg == nullptr)  return false;
 161   }
 162   return reg->is_Region() && reg->in(0) != nullptr && reg->in(0)->is_Root();
 163 }
 164 
 165 void GraphKit::verify_map() const {
 166   if (map() == nullptr)  return;  // null map is OK
 167   assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
 168   assert(!map()->has_exceptions(),    "call add_exception_states_from 1st");
 169   assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
 170 }
 171 
 172 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
 173   assert(ex_map->next_exception() == nullptr, "not already part of a chain");
 174   assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
 175 }
 176 #endif
 177 
 178 //---------------------------stop_and_kill_map---------------------------------
 179 // Set _map to null, signalling a stop to further bytecode execution.
 180 // First smash the current map's control to a constant, to mark it dead.
 181 void GraphKit::stop_and_kill_map() {
 182   SafePointNode* dead_map = stop();
 183   if (dead_map != nullptr) {
 184     dead_map->disconnect_inputs(C); // Mark the map as killed.
 185     assert(dead_map->is_killed(), "must be so marked");
 186   }
 187 }
 188 
 189 
 190 //--------------------------------stopped--------------------------------------
 191 // Tell if _map is null, or control is top.
 192 bool GraphKit::stopped() {
 193   if (map() == nullptr)        return true;
 194   else if (control() == top()) return true;
 195   else                         return false;
 196 }
 197 
 198 
 199 //-----------------------------has_exception_handler----------------------------------
 200 // Tell if this method or any caller method has exception handlers.
 201 bool GraphKit::has_exception_handler() {
 202   for (JVMState* jvmsp = jvms(); jvmsp != nullptr; jvmsp = jvmsp->caller()) {
 203     if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
 204       return true;
 205     }
 206   }
 207   return false;
 208 }
 209 
 210 //------------------------------save_ex_oop------------------------------------
 211 // Save an exception without blowing stack contents or other JVM state.
 212 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
 213   assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
 214   ex_map->add_req(ex_oop);
 215   DEBUG_ONLY(verify_exception_state(ex_map));
 216 }
 217 
 218 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
 219   assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
 220   Node* ex_oop = ex_map->in(ex_map->req()-1);
 221   if (clear_it)  ex_map->del_req(ex_map->req()-1);
 222   return ex_oop;
 223 }
 224 
 225 //-----------------------------saved_ex_oop------------------------------------
 226 // Recover a saved exception from its map.
 227 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
 228   return common_saved_ex_oop(ex_map, false);
 229 }
 230 
 231 //--------------------------clear_saved_ex_oop---------------------------------
 232 // Erase a previously saved exception from its map.
 233 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
 234   return common_saved_ex_oop(ex_map, true);
 235 }
 236 
 237 #ifdef ASSERT
 238 //---------------------------has_saved_ex_oop----------------------------------
 239 // Erase a previously saved exception from its map.
 240 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
 241   return ex_map->req() == ex_map->jvms()->endoff()+1;
 242 }
 243 #endif
 244 
 245 //-------------------------make_exception_state--------------------------------
 246 // Turn the current JVM state into an exception state, appending the ex_oop.
 247 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
 248   sync_jvms();
 249   SafePointNode* ex_map = stop();  // do not manipulate this map any more
 250   set_saved_ex_oop(ex_map, ex_oop);
 251   return ex_map;
 252 }
 253 
 254 
 255 //--------------------------add_exception_state--------------------------------
 256 // Add an exception to my list of exceptions.
 257 void GraphKit::add_exception_state(SafePointNode* ex_map) {
 258   if (ex_map == nullptr || ex_map->control() == top()) {
 259     return;
 260   }
 261 #ifdef ASSERT
 262   verify_exception_state(ex_map);
 263   if (has_exceptions()) {
 264     assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
 265   }
 266 #endif
 267 
 268   // If there is already an exception of exactly this type, merge with it.
 269   // In particular, null-checks and other low-level exceptions common up here.
 270   Node*       ex_oop  = saved_ex_oop(ex_map);
 271   const Type* ex_type = _gvn.type(ex_oop);
 272   if (ex_oop == top()) {
 273     // No action needed.
 274     return;
 275   }
 276   assert(ex_type->isa_instptr(), "exception must be an instance");
 277   for (SafePointNode* e2 = _exceptions; e2 != nullptr; e2 = e2->next_exception()) {
 278     const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
 279     // We check sp also because call bytecodes can generate exceptions
 280     // both before and after arguments are popped!
 281     if (ex_type2 == ex_type
 282         && e2->_jvms->sp() == ex_map->_jvms->sp()) {
 283       combine_exception_states(ex_map, e2);
 284       return;
 285     }
 286   }
 287 
 288   // No pre-existing exception of the same type.  Chain it on the list.
 289   push_exception_state(ex_map);
 290 }
 291 
 292 //-----------------------add_exception_states_from-----------------------------
 293 void GraphKit::add_exception_states_from(JVMState* jvms) {
 294   SafePointNode* ex_map = jvms->map()->next_exception();
 295   if (ex_map != nullptr) {
 296     jvms->map()->set_next_exception(nullptr);
 297     for (SafePointNode* next_map; ex_map != nullptr; ex_map = next_map) {
 298       next_map = ex_map->next_exception();
 299       ex_map->set_next_exception(nullptr);
 300       add_exception_state(ex_map);
 301     }
 302   }
 303 }
 304 
 305 //-----------------------transfer_exceptions_into_jvms-------------------------
 306 JVMState* GraphKit::transfer_exceptions_into_jvms() {
 307   if (map() == nullptr) {
 308     // We need a JVMS to carry the exceptions, but the map has gone away.
 309     // Create a scratch JVMS, cloned from any of the exception states...
 310     if (has_exceptions()) {
 311       _map = _exceptions;
 312       _map = clone_map();
 313       _map->set_next_exception(nullptr);
 314       clear_saved_ex_oop(_map);
 315       DEBUG_ONLY(verify_map());
 316     } else {
 317       // ...or created from scratch
 318       JVMState* jvms = new (C) JVMState(_method, nullptr);
 319       jvms->set_bci(_bci);
 320       jvms->set_sp(_sp);
 321       jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms));
 322       set_jvms(jvms);
 323       for (uint i = 0; i < map()->req(); i++)  map()->init_req(i, top());
 324       set_all_memory(top());
 325       while (map()->req() < jvms->endoff())  map()->add_req(top());
 326     }
 327     // (This is a kludge, in case you didn't notice.)
 328     set_control(top());
 329   }
 330   JVMState* jvms = sync_jvms();
 331   assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
 332   jvms->map()->set_next_exception(_exceptions);
 333   _exceptions = nullptr;   // done with this set of exceptions
 334   return jvms;
 335 }
 336 
 337 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
 338   assert(is_hidden_merge(dstphi), "must be a special merge node");
 339   assert(is_hidden_merge(srcphi), "must be a special merge node");
 340   uint limit = srcphi->req();
 341   for (uint i = PhiNode::Input; i < limit; i++) {
 342     dstphi->add_req(srcphi->in(i));
 343   }
 344 }
 345 static inline void add_one_req(Node* dstphi, Node* src) {
 346   assert(is_hidden_merge(dstphi), "must be a special merge node");
 347   assert(!is_hidden_merge(src), "must not be a special merge node");
 348   dstphi->add_req(src);
 349 }
 350 
 351 //-----------------------combine_exception_states------------------------------
 352 // This helper function combines exception states by building phis on a
 353 // specially marked state-merging region.  These regions and phis are
 354 // untransformed, and can build up gradually.  The region is marked by
 355 // having a control input of its exception map, rather than null.  Such
 356 // regions do not appear except in this function, and in use_exception_state.
 357 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
 358   if (failing_internal()) {
 359     return;  // dying anyway...
 360   }
 361   JVMState* ex_jvms = ex_map->_jvms;
 362   assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
 363   assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
 364   // TODO 8325632 Re-enable
 365   // assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
 366   assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
 367   assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
 368   assert(ex_map->req() == phi_map->req(), "matching maps");
 369   uint tos = ex_jvms->stkoff() + ex_jvms->sp();
 370   Node*         hidden_merge_mark = root();
 371   Node*         region  = phi_map->control();
 372   MergeMemNode* phi_mem = phi_map->merged_memory();
 373   MergeMemNode* ex_mem  = ex_map->merged_memory();
 374   if (region->in(0) != hidden_merge_mark) {
 375     // The control input is not (yet) a specially-marked region in phi_map.
 376     // Make it so, and build some phis.
 377     region = new RegionNode(2);
 378     _gvn.set_type(region, Type::CONTROL);
 379     region->set_req(0, hidden_merge_mark);  // marks an internal ex-state
 380     region->init_req(1, phi_map->control());
 381     phi_map->set_control(region);
 382     Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
 383     record_for_igvn(io_phi);
 384     _gvn.set_type(io_phi, Type::ABIO);
 385     phi_map->set_i_o(io_phi);
 386     for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
 387       Node* m = mms.memory();
 388       Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
 389       record_for_igvn(m_phi);
 390       _gvn.set_type(m_phi, Type::MEMORY);
 391       mms.set_memory(m_phi);
 392     }
 393   }
 394 
 395   // Either or both of phi_map and ex_map might already be converted into phis.
 396   Node* ex_control = ex_map->control();
 397   // if there is special marking on ex_map also, we add multiple edges from src
 398   bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
 399   // how wide was the destination phi_map, originally?
 400   uint orig_width = region->req();
 401 
 402   if (add_multiple) {
 403     add_n_reqs(region, ex_control);
 404     add_n_reqs(phi_map->i_o(), ex_map->i_o());
 405   } else {
 406     // ex_map has no merges, so we just add single edges everywhere
 407     add_one_req(region, ex_control);
 408     add_one_req(phi_map->i_o(), ex_map->i_o());
 409   }
 410   for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
 411     if (mms.is_empty()) {
 412       // get a copy of the base memory, and patch some inputs into it
 413       const TypePtr* adr_type = mms.adr_type(C);
 414       Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
 415       assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
 416       mms.set_memory(phi);
 417       // Prepare to append interesting stuff onto the newly sliced phi:
 418       while (phi->req() > orig_width)  phi->del_req(phi->req()-1);
 419     }
 420     // Append stuff from ex_map:
 421     if (add_multiple) {
 422       add_n_reqs(mms.memory(), mms.memory2());
 423     } else {
 424       add_one_req(mms.memory(), mms.memory2());
 425     }
 426   }
 427   uint limit = ex_map->req();
 428   for (uint i = TypeFunc::Parms; i < limit; i++) {
 429     // Skip everything in the JVMS after tos.  (The ex_oop follows.)
 430     if (i == tos)  i = ex_jvms->monoff();
 431     Node* src = ex_map->in(i);
 432     Node* dst = phi_map->in(i);
 433     if (src != dst) {
 434       PhiNode* phi;
 435       if (dst->in(0) != region) {
 436         dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
 437         record_for_igvn(phi);
 438         _gvn.set_type(phi, phi->type());
 439         phi_map->set_req(i, dst);
 440         // Prepare to append interesting stuff onto the new phi:
 441         while (dst->req() > orig_width)  dst->del_req(dst->req()-1);
 442       } else {
 443         assert(dst->is_Phi(), "nobody else uses a hidden region");
 444         phi = dst->as_Phi();
 445       }
 446       if (add_multiple && src->in(0) == ex_control) {
 447         // Both are phis.
 448         add_n_reqs(dst, src);
 449       } else {
 450         while (dst->req() < region->req())  add_one_req(dst, src);
 451       }
 452       const Type* srctype = _gvn.type(src);
 453       if (phi->type() != srctype) {
 454         const Type* dsttype = phi->type()->meet_speculative(srctype);
 455         if (phi->type() != dsttype) {
 456           phi->set_type(dsttype);
 457           _gvn.set_type(phi, dsttype);
 458         }
 459       }
 460     }
 461   }
 462   phi_map->merge_replaced_nodes_with(ex_map);
 463 }
 464 
 465 //--------------------------use_exception_state--------------------------------
 466 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
 467   if (failing_internal()) { stop(); return top(); }
 468   Node* region = phi_map->control();
 469   Node* hidden_merge_mark = root();
 470   assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
 471   Node* ex_oop = clear_saved_ex_oop(phi_map);
 472   if (region->in(0) == hidden_merge_mark) {
 473     // Special marking for internal ex-states.  Process the phis now.
 474     region->set_req(0, region);  // now it's an ordinary region
 475     set_jvms(phi_map->jvms());   // ...so now we can use it as a map
 476     // Note: Setting the jvms also sets the bci and sp.
 477     set_control(_gvn.transform(region));
 478     uint tos = jvms()->stkoff() + sp();
 479     for (uint i = 1; i < tos; i++) {
 480       Node* x = phi_map->in(i);
 481       if (x->in(0) == region) {
 482         assert(x->is_Phi(), "expected a special phi");
 483         phi_map->set_req(i, _gvn.transform(x));
 484       }
 485     }
 486     for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
 487       Node* x = mms.memory();
 488       if (x->in(0) == region) {
 489         assert(x->is_Phi(), "nobody else uses a hidden region");
 490         mms.set_memory(_gvn.transform(x));
 491       }
 492     }
 493     if (ex_oop->in(0) == region) {
 494       assert(ex_oop->is_Phi(), "expected a special phi");
 495       ex_oop = _gvn.transform(ex_oop);
 496     }
 497   } else {
 498     set_jvms(phi_map->jvms());
 499   }
 500 
 501   assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
 502   assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
 503   return ex_oop;
 504 }
 505 
 506 //---------------------------------java_bc-------------------------------------
 507 Bytecodes::Code GraphKit::java_bc() const {
 508   ciMethod* method = this->method();
 509   int       bci    = this->bci();
 510   if (method != nullptr && bci != InvocationEntryBci)
 511     return method->java_code_at_bci(bci);
 512   else
 513     return Bytecodes::_illegal;
 514 }
 515 
 516 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
 517                                                           bool must_throw) {
 518     // if the exception capability is set, then we will generate code
 519     // to check the JavaThread.should_post_on_exceptions flag to see
 520     // if we actually need to report exception events (for this
 521     // thread).  If we don't need to report exception events, we will
 522     // take the normal fast path provided by add_exception_events.  If
 523     // exception event reporting is enabled for this thread, we will
 524     // take the uncommon_trap in the BuildCutout below.
 525 
 526     // first must access the should_post_on_exceptions_flag in this thread's JavaThread
 527     Node* jthread = _gvn.transform(new ThreadLocalNode());
 528     Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
 529     Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
 530 
 531     // Test the should_post_on_exceptions_flag vs. 0
 532     Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) );
 533     Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) );
 534 
 535     // Branch to slow_path if should_post_on_exceptions_flag was true
 536     { BuildCutout unless(this, tst, PROB_MAX);
 537       // Do not try anything fancy if we're notifying the VM on every throw.
 538       // Cf. case Bytecodes::_athrow in parse2.cpp.
 539       uncommon_trap(reason, Deoptimization::Action_none,
 540                     (ciKlass*)nullptr, (char*)nullptr, must_throw);
 541     }
 542 
 543 }
 544 
 545 //------------------------------builtin_throw----------------------------------
 546 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason) {
 547   builtin_throw(reason, builtin_throw_exception(reason), /*allow_too_many_traps*/ true);
 548 }
 549 
 550 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason,
 551                              ciInstance* ex_obj,
 552                              bool allow_too_many_traps) {
 553   // If this throw happens frequently, an uncommon trap might cause
 554   // a performance pothole.  If there is a local exception handler,
 555   // and if this particular bytecode appears to be deoptimizing often,
 556   // let us handle the throw inline, with a preconstructed instance.
 557   // Note:   If the deopt count has blown up, the uncommon trap
 558   // runtime is going to flush this nmethod, not matter what.
 559   if (is_builtin_throw_hot(reason)) {
 560     if (method()->can_omit_stack_trace() && ex_obj != nullptr) {
 561       // If the throw is local, we use a pre-existing instance and
 562       // punt on the backtrace.  This would lead to a missing backtrace
 563       // (a repeat of 4292742) if the backtrace object is ever asked
 564       // for its backtrace.
 565       // Fixing this remaining case of 4292742 requires some flavor of
 566       // escape analysis.  Leave that for the future.
 567       if (env()->jvmti_can_post_on_exceptions()) {
 568         // check if we must post exception events, take uncommon trap if so
 569         uncommon_trap_if_should_post_on_exceptions(reason, true /*must_throw*/);
 570         // here if should_post_on_exceptions is false
 571         // continue on with the normal codegen
 572       }
 573 
 574       // Cheat with a preallocated exception object.
 575       if (C->log() != nullptr)
 576         C->log()->elem("hot_throw preallocated='1' reason='%s'",
 577                        Deoptimization::trap_reason_name(reason));
 578       const TypeInstPtr* ex_con  = TypeInstPtr::make(ex_obj);
 579       Node*              ex_node = _gvn.transform(ConNode::make(ex_con));
 580 
 581       // Clear the detail message of the preallocated exception object.
 582       // Weblogic sometimes mutates the detail message of exceptions
 583       // using reflection.
 584       int offset = java_lang_Throwable::get_detailMessage_offset();
 585       const TypePtr* adr_typ = ex_con->add_offset(offset);
 586 
 587       Node *adr = basic_plus_adr(ex_node, ex_node, offset);
 588       const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
 589       Node *store = access_store_at(ex_node, adr, adr_typ, null(), val_type, T_OBJECT, IN_HEAP);
 590 
 591       if (!method()->has_exception_handlers()) {
 592         // We don't need to preserve the stack if there's no handler as the entire frame is going to be popped anyway.
 593         // This prevents issues with exception handling and late inlining.
 594         set_sp(0);
 595         clean_stack(0);
 596       }
 597 
 598       add_exception_state(make_exception_state(ex_node));
 599       return;
 600     } else if (builtin_throw_too_many_traps(reason, ex_obj)) {
 601       // We cannot afford to take too many traps here. Suffer in the interpreter instead.
 602       assert(allow_too_many_traps, "not allowed");
 603       if (C->log() != nullptr) {
 604         C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
 605                        Deoptimization::trap_reason_name(reason),
 606                        C->trap_count(reason));
 607       }
 608       uncommon_trap(reason, Deoptimization::Action_none,
 609                     (ciKlass*) nullptr, (char*) nullptr,
 610                     true /*must_throw*/);
 611       return;
 612     }
 613   }
 614 
 615   // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
 616   // It won't be much cheaper than bailing to the interp., since we'll
 617   // have to pass up all the debug-info, and the runtime will have to
 618   // create the stack trace.
 619 
 620   // Usual case:  Bail to interpreter.
 621   // Reserve the right to recompile if we haven't seen anything yet.
 622 
 623   // "must_throw" prunes the JVM state to include only the stack, if there
 624   // are no local exception handlers.  This should cut down on register
 625   // allocation time and code size, by drastically reducing the number
 626   // of in-edges on the call to the uncommon trap.
 627   uncommon_trap(reason, Deoptimization::Action_maybe_recompile,
 628                 (ciKlass*) nullptr, (char*) nullptr,
 629                 true /*must_throw*/);
 630 }
 631 
 632 bool GraphKit::is_builtin_throw_hot(Deoptimization::DeoptReason reason) {
 633   // If this particular condition has not yet happened at this
 634   // bytecode, then use the uncommon trap mechanism, and allow for
 635   // a future recompilation if several traps occur here.
 636   // If the throw is hot, try to use a more complicated inline mechanism
 637   // which keeps execution inside the compiled code.
 638   if (ProfileTraps) {
 639     if (too_many_traps(reason)) {
 640       return true;
 641     }
 642     // (If there is no MDO at all, assume it is early in
 643     // execution, and that any deopts are part of the
 644     // startup transient, and don't need to be remembered.)
 645 
 646     // Also, if there is a local exception handler, treat all throws
 647     // as hot if there has been at least one in this method.
 648     if (C->trap_count(reason) != 0 &&
 649         method()->method_data()->trap_count(reason) != 0 &&
 650         has_exception_handler()) {
 651       return true;
 652     }
 653   }
 654   return false;
 655 }
 656 
 657 bool GraphKit::builtin_throw_too_many_traps(Deoptimization::DeoptReason reason,
 658                                             ciInstance* ex_obj) {
 659   if (is_builtin_throw_hot(reason)) {
 660     if (method()->can_omit_stack_trace() && ex_obj != nullptr) {
 661       return false; // no traps; throws preallocated exception instead
 662     }
 663     ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : nullptr;
 664     if (method()->method_data()->trap_recompiled_at(bci(), m) ||
 665         C->too_many_traps(reason)) {
 666       return true;
 667     }
 668   }
 669   return false;
 670 }
 671 
 672 ciInstance* GraphKit::builtin_throw_exception(Deoptimization::DeoptReason reason) const {
 673   // Preallocated exception objects to use when we don't need the backtrace.
 674   switch (reason) {
 675   case Deoptimization::Reason_null_check:
 676     return env()->NullPointerException_instance();
 677   case Deoptimization::Reason_div0_check:
 678     return env()->ArithmeticException_instance();
 679   case Deoptimization::Reason_range_check:
 680     return env()->ArrayIndexOutOfBoundsException_instance();
 681   case Deoptimization::Reason_class_check:
 682     return env()->ClassCastException_instance();
 683   case Deoptimization::Reason_array_check:
 684     return env()->ArrayStoreException_instance();
 685   default:
 686     return nullptr;
 687   }
 688 }
 689 
 690 //----------------------------PreserveJVMState---------------------------------
 691 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
 692   DEBUG_ONLY(kit->verify_map());
 693   _kit    = kit;
 694   _map    = kit->map();   // preserve the map
 695   _sp     = kit->sp();
 696   kit->set_map(clone_map ? kit->clone_map() : nullptr);
 697 #ifdef ASSERT
 698   _bci    = kit->bci();
 699   Parse* parser = kit->is_Parse();
 700   int block = (parser == nullptr || parser->block() == nullptr) ? -1 : parser->block()->rpo();
 701   _block  = block;
 702 #endif
 703 }
 704 PreserveJVMState::~PreserveJVMState() {
 705   GraphKit* kit = _kit;
 706 #ifdef ASSERT
 707   assert(kit->bci() == _bci, "bci must not shift");
 708   Parse* parser = kit->is_Parse();
 709   int block = (parser == nullptr || parser->block() == nullptr) ? -1 : parser->block()->rpo();
 710   assert(block == _block,    "block must not shift");
 711 #endif
 712   kit->set_map(_map);
 713   kit->set_sp(_sp);
 714 }
 715 
 716 
 717 //-----------------------------BuildCutout-------------------------------------
 718 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
 719   : PreserveJVMState(kit)
 720 {
 721   assert(p->is_Con() || p->is_Bool(), "test must be a bool");
 722   SafePointNode* outer_map = _map;   // preserved map is caller's
 723   SafePointNode* inner_map = kit->map();
 724   IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
 725   outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) ));
 726   inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) ));
 727 }
 728 BuildCutout::~BuildCutout() {
 729   GraphKit* kit = _kit;
 730   assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
 731 }
 732 
 733 //---------------------------PreserveReexecuteState----------------------------
 734 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
 735   assert(!kit->stopped(), "must call stopped() before");
 736   _kit    =    kit;
 737   _sp     =    kit->sp();
 738   _reexecute = kit->jvms()->_reexecute;
 739 }
 740 PreserveReexecuteState::~PreserveReexecuteState() {
 741   if (_kit->stopped()) return;
 742   _kit->jvms()->_reexecute = _reexecute;
 743   _kit->set_sp(_sp);
 744 }
 745 
 746 //------------------------------clone_map--------------------------------------
 747 // Implementation of PreserveJVMState
 748 //
 749 // Only clone_map(...) here. If this function is only used in the
 750 // PreserveJVMState class we may want to get rid of this extra
 751 // function eventually and do it all there.
 752 
 753 SafePointNode* GraphKit::clone_map() {
 754   if (map() == nullptr)  return nullptr;
 755 
 756   // Clone the memory edge first
 757   Node* mem = MergeMemNode::make(map()->memory());
 758   gvn().set_type_bottom(mem);
 759 
 760   SafePointNode *clonemap = (SafePointNode*)map()->clone();
 761   JVMState* jvms = this->jvms();
 762   JVMState* clonejvms = jvms->clone_shallow(C);
 763   clonemap->set_memory(mem);
 764   clonemap->set_jvms(clonejvms);
 765   clonejvms->set_map(clonemap);
 766   record_for_igvn(clonemap);
 767   gvn().set_type_bottom(clonemap);
 768   return clonemap;
 769 }
 770 
 771 //-----------------------------destruct_map_clone------------------------------
 772 //
 773 // Order of destruct is important to increase the likelyhood that memory can be re-used. We need
 774 // to destruct/free/delete in the exact opposite order as clone_map().
 775 void GraphKit::destruct_map_clone(SafePointNode* sfp) {
 776   if (sfp == nullptr) return;
 777 
 778   Node* mem = sfp->memory();
 779   JVMState* jvms = sfp->jvms();
 780 
 781   if (jvms != nullptr) {
 782     delete jvms;
 783   }
 784 
 785   remove_for_igvn(sfp);
 786   gvn().clear_type(sfp);
 787   sfp->destruct(&_gvn);
 788 
 789   if (mem != nullptr) {
 790     gvn().clear_type(mem);
 791     mem->destruct(&_gvn);
 792   }
 793 }
 794 
 795 //-----------------------------set_map_clone-----------------------------------
 796 void GraphKit::set_map_clone(SafePointNode* m) {
 797   _map = m;
 798   _map = clone_map();
 799   _map->set_next_exception(nullptr);
 800   DEBUG_ONLY(verify_map());
 801 }
 802 
 803 
 804 //----------------------------kill_dead_locals---------------------------------
 805 // Detect any locals which are known to be dead, and force them to top.
 806 void GraphKit::kill_dead_locals() {
 807   // Consult the liveness information for the locals.  If any
 808   // of them are unused, then they can be replaced by top().  This
 809   // should help register allocation time and cut down on the size
 810   // of the deoptimization information.
 811 
 812   // This call is made from many of the bytecode handling
 813   // subroutines called from the Big Switch in do_one_bytecode.
 814   // Every bytecode which might include a slow path is responsible
 815   // for killing its dead locals.  The more consistent we
 816   // are about killing deads, the fewer useless phis will be
 817   // constructed for them at various merge points.
 818 
 819   // bci can be -1 (InvocationEntryBci).  We return the entry
 820   // liveness for the method.
 821 
 822   if (method() == nullptr || method()->code_size() == 0) {
 823     // We are building a graph for a call to a native method.
 824     // All locals are live.
 825     return;
 826   }
 827 
 828   ResourceMark rm;
 829 
 830   // Consult the liveness information for the locals.  If any
 831   // of them are unused, then they can be replaced by top().  This
 832   // should help register allocation time and cut down on the size
 833   // of the deoptimization information.
 834   MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
 835 
 836   int len = (int)live_locals.size();
 837   assert(len <= jvms()->loc_size(), "too many live locals");
 838   for (int local = 0; local < len; local++) {
 839     if (!live_locals.at(local)) {
 840       set_local(local, top());
 841     }
 842   }
 843 }
 844 
 845 #ifdef ASSERT
 846 //-------------------------dead_locals_are_killed------------------------------
 847 // Return true if all dead locals are set to top in the map.
 848 // Used to assert "clean" debug info at various points.
 849 bool GraphKit::dead_locals_are_killed() {
 850   if (method() == nullptr || method()->code_size() == 0) {
 851     // No locals need to be dead, so all is as it should be.
 852     return true;
 853   }
 854 
 855   // Make sure somebody called kill_dead_locals upstream.
 856   ResourceMark rm;
 857   for (JVMState* jvms = this->jvms(); jvms != nullptr; jvms = jvms->caller()) {
 858     if (jvms->loc_size() == 0)  continue;  // no locals to consult
 859     SafePointNode* map = jvms->map();
 860     ciMethod* method = jvms->method();
 861     int       bci    = jvms->bci();
 862     if (jvms == this->jvms()) {
 863       bci = this->bci();  // it might not yet be synched
 864     }
 865     MethodLivenessResult live_locals = method->liveness_at_bci(bci);
 866     int len = (int)live_locals.size();
 867     if (!live_locals.is_valid() || len == 0)
 868       // This method is trivial, or is poisoned by a breakpoint.
 869       return true;
 870     assert(len == jvms->loc_size(), "live map consistent with locals map");
 871     for (int local = 0; local < len; local++) {
 872       if (!live_locals.at(local) && map->local(jvms, local) != top()) {
 873         if (PrintMiscellaneous && (Verbose || WizardMode)) {
 874           tty->print_cr("Zombie local %d: ", local);
 875           jvms->dump();
 876         }
 877         return false;
 878       }
 879     }
 880   }
 881   return true;
 882 }
 883 
 884 #endif //ASSERT
 885 
 886 // Helper function for enforcing certain bytecodes to reexecute if deoptimization happens.
 887 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
 888   ciMethod* cur_method = jvms->method();
 889   int       cur_bci   = jvms->bci();
 890   if (cur_method != nullptr && cur_bci != InvocationEntryBci) {
 891     Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
 892     return Interpreter::bytecode_should_reexecute(code) ||
 893            (is_anewarray && (code == Bytecodes::_multianewarray));
 894     // Reexecute _multianewarray bytecode which was replaced with
 895     // sequence of [a]newarray. See Parse::do_multianewarray().
 896     //
 897     // Note: interpreter should not have it set since this optimization
 898     // is limited by dimensions and guarded by flag so in some cases
 899     // multianewarray() runtime calls will be generated and
 900     // the bytecode should not be reexecutes (stack will not be reset).
 901   } else {
 902     return false;
 903   }
 904 }
 905 
 906 // Helper function for adding JVMState and debug information to node
 907 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
 908   // Add the safepoint edges to the call (or other safepoint).
 909 
 910   // Make sure dead locals are set to top.  This
 911   // should help register allocation time and cut down on the size
 912   // of the deoptimization information.
 913   assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
 914 
 915   // Walk the inline list to fill in the correct set of JVMState's
 916   // Also fill in the associated edges for each JVMState.
 917 
 918   // If the bytecode needs to be reexecuted we need to put
 919   // the arguments back on the stack.
 920   const bool should_reexecute = jvms()->should_reexecute();
 921   JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
 922 
 923   // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
 924   // undefined if the bci is different.  This is normal for Parse but it
 925   // should not happen for LibraryCallKit because only one bci is processed.
 926   assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
 927          "in LibraryCallKit the reexecute bit should not change");
 928 
 929   // If we are guaranteed to throw, we can prune everything but the
 930   // input to the current bytecode.
 931   bool can_prune_locals = false;
 932   uint stack_slots_not_pruned = 0;
 933   int inputs = 0, depth = 0;
 934   if (must_throw) {
 935     assert(method() == youngest_jvms->method(), "sanity");
 936     if (compute_stack_effects(inputs, depth)) {
 937       can_prune_locals = true;
 938       stack_slots_not_pruned = inputs;
 939     }
 940   }
 941 
 942   if (env()->should_retain_local_variables()) {
 943     // At any safepoint, this method can get breakpointed, which would
 944     // then require an immediate deoptimization.
 945     can_prune_locals = false;  // do not prune locals
 946     stack_slots_not_pruned = 0;
 947   }
 948 
 949   // do not scribble on the input jvms
 950   JVMState* out_jvms = youngest_jvms->clone_deep(C);
 951   call->set_jvms(out_jvms); // Start jvms list for call node
 952 
 953   // For a known set of bytecodes, the interpreter should reexecute them if
 954   // deoptimization happens. We set the reexecute state for them here
 955   if (out_jvms->is_reexecute_undefined() && //don't change if already specified
 956       should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
 957 #ifdef ASSERT
 958     int inputs = 0, not_used; // initialized by GraphKit::compute_stack_effects()
 959     assert(method() == youngest_jvms->method(), "sanity");
 960     assert(compute_stack_effects(inputs, not_used), "unknown bytecode: %s", Bytecodes::name(java_bc()));
 961     // TODO 8371125
 962     // assert(out_jvms->sp() >= (uint)inputs, "not enough operands for reexecution");
 963 #endif // ASSERT
 964     out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
 965   }
 966 
 967   // Presize the call:
 968   DEBUG_ONLY(uint non_debug_edges = call->req());
 969   call->add_req_batch(top(), youngest_jvms->debug_depth());
 970   assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
 971 
 972   // Set up edges so that the call looks like this:
 973   //  Call [state:] ctl io mem fptr retadr
 974   //       [parms:] parm0 ... parmN
 975   //       [root:]  loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
 976   //    [...mid:]   loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
 977   //       [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
 978   // Note that caller debug info precedes callee debug info.
 979 
 980   // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
 981   uint debug_ptr = call->req();
 982 
 983   // Loop over the map input edges associated with jvms, add them
 984   // to the call node, & reset all offsets to match call node array.
 985 
 986   JVMState* callee_jvms = nullptr;
 987   for (JVMState* in_jvms = youngest_jvms; in_jvms != nullptr; ) {
 988     uint debug_end   = debug_ptr;
 989     uint debug_start = debug_ptr - in_jvms->debug_size();
 990     debug_ptr = debug_start;  // back up the ptr
 991 
 992     uint p = debug_start;  // walks forward in [debug_start, debug_end)
 993     uint j, k, l;
 994     SafePointNode* in_map = in_jvms->map();
 995     out_jvms->set_map(call);
 996 
 997     if (can_prune_locals) {
 998       assert(in_jvms->method() == out_jvms->method(), "sanity");
 999       // If the current throw can reach an exception handler in this JVMS,
1000       // then we must keep everything live that can reach that handler.
1001       // As a quick and dirty approximation, we look for any handlers at all.
1002       if (in_jvms->method()->has_exception_handlers()) {
1003         can_prune_locals = false;
1004       }
1005     }
1006 
1007     // Add the Locals
1008     k = in_jvms->locoff();
1009     l = in_jvms->loc_size();
1010     out_jvms->set_locoff(p);
1011     if (!can_prune_locals) {
1012       for (j = 0; j < l; j++) {
1013         call->set_req(p++, in_map->in(k + j));
1014       }
1015     } else {
1016       p += l;  // already set to top above by add_req_batch
1017     }
1018 
1019     // Add the Expression Stack
1020     k = in_jvms->stkoff();
1021     l = in_jvms->sp();
1022     out_jvms->set_stkoff(p);
1023     if (!can_prune_locals) {
1024       for (j = 0; j < l; j++) {
1025         call->set_req(p++, in_map->in(k + j));
1026       }
1027     } else if (can_prune_locals && stack_slots_not_pruned != 0) {
1028       // Divide stack into {S0,...,S1}, where S0 is set to top.
1029       uint s1 = stack_slots_not_pruned;
1030       stack_slots_not_pruned = 0;  // for next iteration
1031       if (s1 > l)  s1 = l;
1032       uint s0 = l - s1;
1033       p += s0;  // skip the tops preinstalled by add_req_batch
1034       for (j = s0; j < l; j++)
1035         call->set_req(p++, in_map->in(k+j));
1036     } else {
1037       p += l;  // already set to top above by add_req_batch
1038     }
1039 
1040     // Add the Monitors
1041     k = in_jvms->monoff();
1042     l = in_jvms->mon_size();
1043     out_jvms->set_monoff(p);
1044     for (j = 0; j < l; j++)
1045       call->set_req(p++, in_map->in(k+j));
1046 
1047     // Copy any scalar object fields.
1048     k = in_jvms->scloff();
1049     l = in_jvms->scl_size();
1050     out_jvms->set_scloff(p);
1051     for (j = 0; j < l; j++)
1052       call->set_req(p++, in_map->in(k+j));
1053 
1054     // Finish the new jvms.
1055     out_jvms->set_endoff(p);
1056 
1057     assert(out_jvms->endoff()     == debug_end,             "fill ptr must match");
1058     assert(out_jvms->depth()      == in_jvms->depth(),      "depth must match");
1059     assert(out_jvms->loc_size()   == in_jvms->loc_size(),   "size must match");
1060     assert(out_jvms->mon_size()   == in_jvms->mon_size(),   "size must match");
1061     assert(out_jvms->scl_size()   == in_jvms->scl_size(),   "size must match");
1062     assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
1063 
1064     // Update the two tail pointers in parallel.
1065     callee_jvms = out_jvms;
1066     out_jvms = out_jvms->caller();
1067     in_jvms  = in_jvms->caller();
1068   }
1069 
1070   assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
1071 
1072   // Test the correctness of JVMState::debug_xxx accessors:
1073   assert(call->jvms()->debug_start() == non_debug_edges, "");
1074   assert(call->jvms()->debug_end()   == call->req(), "");
1075   assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1076 }
1077 
1078 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1079   Bytecodes::Code code = java_bc();
1080   if (code == Bytecodes::_wide) {
1081     code = method()->java_code_at_bci(bci() + 1);
1082   }
1083 
1084   if (code != Bytecodes::_illegal) {
1085     depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1086   }
1087 
1088   auto rsize = [&]() {
1089     assert(code != Bytecodes::_illegal, "code is illegal!");
1090     BasicType rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1091     return (rtype < T_CONFLICT) ? type2size[rtype] : 0;
1092   };
1093 
1094   switch (code) {
1095   case Bytecodes::_illegal:
1096     return false;
1097 
1098   case Bytecodes::_ldc:
1099   case Bytecodes::_ldc_w:
1100   case Bytecodes::_ldc2_w:
1101     inputs = 0;
1102     break;
1103 
1104   case Bytecodes::_dup:         inputs = 1;  break;
1105   case Bytecodes::_dup_x1:      inputs = 2;  break;
1106   case Bytecodes::_dup_x2:      inputs = 3;  break;
1107   case Bytecodes::_dup2:        inputs = 2;  break;
1108   case Bytecodes::_dup2_x1:     inputs = 3;  break;
1109   case Bytecodes::_dup2_x2:     inputs = 4;  break;
1110   case Bytecodes::_swap:        inputs = 2;  break;
1111   case Bytecodes::_arraylength: inputs = 1;  break;
1112 
1113   case Bytecodes::_getstatic:
1114   case Bytecodes::_putstatic:
1115   case Bytecodes::_getfield:
1116   case Bytecodes::_putfield:
1117     {
1118       bool ignored_will_link;
1119       ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1120       int      size  = field->type()->size();
1121       bool is_get = (depth >= 0), is_static = (depth & 1);
1122       inputs = (is_static ? 0 : 1);
1123       if (is_get) {
1124         depth = size - inputs;
1125       } else {
1126         inputs += size;        // putxxx pops the value from the stack
1127         depth = - inputs;
1128       }
1129     }
1130     break;
1131 
1132   case Bytecodes::_invokevirtual:
1133   case Bytecodes::_invokespecial:
1134   case Bytecodes::_invokestatic:
1135   case Bytecodes::_invokedynamic:
1136   case Bytecodes::_invokeinterface:
1137     {
1138       bool ignored_will_link;
1139       ciSignature* declared_signature = nullptr;
1140       ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1141       assert(declared_signature != nullptr, "cannot be null");
1142       inputs   = declared_signature->arg_size_for_bc(code);
1143       int size = declared_signature->return_type()->size();
1144       depth = size - inputs;
1145     }
1146     break;
1147 
1148   case Bytecodes::_multianewarray:
1149     {
1150       ciBytecodeStream iter(method());
1151       iter.reset_to_bci(bci());
1152       iter.next();
1153       inputs = iter.get_dimensions();
1154       assert(rsize() == 1, "");
1155       depth = 1 - inputs;
1156     }
1157     break;
1158 
1159   case Bytecodes::_ireturn:
1160   case Bytecodes::_lreturn:
1161   case Bytecodes::_freturn:
1162   case Bytecodes::_dreturn:
1163   case Bytecodes::_areturn:
1164     assert(rsize() == -depth, "");
1165     inputs = -depth;
1166     break;
1167 
1168   case Bytecodes::_jsr:
1169   case Bytecodes::_jsr_w:
1170     inputs = 0;
1171     depth  = 1;                  // S.B. depth=1, not zero
1172     break;
1173 
1174   default:
1175     // bytecode produces a typed result
1176     inputs = rsize() - depth;
1177     assert(inputs >= 0, "");
1178     break;
1179   }
1180 
1181 #ifdef ASSERT
1182   // spot check
1183   int outputs = depth + inputs;
1184   assert(outputs >= 0, "sanity");
1185   switch (code) {
1186   case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1187   case Bytecodes::_athrow:    assert(inputs == 1 && outputs == 0, ""); break;
1188   case Bytecodes::_aload_0:   assert(inputs == 0 && outputs == 1, ""); break;
1189   case Bytecodes::_return:    assert(inputs == 0 && outputs == 0, ""); break;
1190   case Bytecodes::_drem:      assert(inputs == 4 && outputs == 2, ""); break;
1191   default:                    break;
1192   }
1193 #endif //ASSERT
1194 
1195   return true;
1196 }
1197 
1198 
1199 
1200 //------------------------------basic_plus_adr---------------------------------
1201 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1202   // short-circuit a common case
1203   if (offset == intcon(0))  return ptr;
1204   return _gvn.transform( new AddPNode(base, ptr, offset) );
1205 }
1206 
1207 Node* GraphKit::ConvI2L(Node* offset) {
1208   // short-circuit a common case
1209   jint offset_con = find_int_con(offset, Type::OffsetBot);
1210   if (offset_con != Type::OffsetBot) {
1211     return longcon((jlong) offset_con);
1212   }
1213   return _gvn.transform( new ConvI2LNode(offset));
1214 }
1215 
1216 Node* GraphKit::ConvI2UL(Node* offset) {
1217   juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1218   if (offset_con != (juint) Type::OffsetBot) {
1219     return longcon((julong) offset_con);
1220   }
1221   Node* conv = _gvn.transform( new ConvI2LNode(offset));
1222   Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1223   return _gvn.transform( new AndLNode(conv, mask) );
1224 }
1225 
1226 Node* GraphKit::ConvL2I(Node* offset) {
1227   // short-circuit a common case
1228   jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1229   if (offset_con != (jlong)Type::OffsetBot) {
1230     return intcon((int) offset_con);
1231   }
1232   return _gvn.transform( new ConvL2INode(offset));
1233 }
1234 
1235 //-------------------------load_object_klass-----------------------------------
1236 Node* GraphKit::load_object_klass(Node* obj) {
1237   // Special-case a fresh allocation to avoid building nodes:
1238   Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1239   if (akls != nullptr)  return akls;
1240   Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1241   return _gvn.transform(LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
1242 }
1243 
1244 //-------------------------load_array_length-----------------------------------
1245 Node* GraphKit::load_array_length(Node* array) {
1246   // Special-case a fresh allocation to avoid building nodes:
1247   AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array);
1248   Node *alen;
1249   if (alloc == nullptr) {
1250     Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1251     alen = _gvn.transform( new LoadRangeNode(nullptr, immutable_memory(), r_adr, TypeInt::POS));
1252   } else {
1253     alen = array_ideal_length(alloc, _gvn.type(array)->is_oopptr(), false);
1254   }
1255   return alen;
1256 }
1257 
1258 Node* GraphKit::array_ideal_length(AllocateArrayNode* alloc,
1259                                    const TypeOopPtr* oop_type,
1260                                    bool replace_length_in_map) {
1261   Node* length = alloc->Ideal_length();
1262   if (replace_length_in_map == false || map()->find_edge(length) >= 0) {
1263     Node* ccast = alloc->make_ideal_length(oop_type, &_gvn);
1264     if (ccast != length) {
1265       // do not transform ccast here, it might convert to top node for
1266       // negative array length and break assumptions in parsing stage.
1267       _gvn.set_type_bottom(ccast);
1268       record_for_igvn(ccast);
1269       if (replace_length_in_map) {
1270         replace_in_map(length, ccast);
1271       }
1272       return ccast;
1273     }
1274   }
1275   return length;
1276 }
1277 
1278 //------------------------------do_null_check----------------------------------
1279 // Helper function to do a null pointer check.  Returned value is
1280 // the incoming address with null casted away.  You are allowed to use the
1281 // not-null value only if you are control dependent on the test.
1282 #ifndef PRODUCT
1283 extern uint explicit_null_checks_inserted,
1284             explicit_null_checks_elided;
1285 #endif
1286 Node* GraphKit::null_check_common(Node* value, BasicType type,
1287                                   // optional arguments for variations:
1288                                   bool assert_null,
1289                                   Node* *null_control,
1290                                   bool speculative,
1291                                   bool null_marker_check) {
1292   assert(!assert_null || null_control == nullptr, "not both at once");
1293   if (stopped())  return top();
1294   NOT_PRODUCT(explicit_null_checks_inserted++);
1295 
1296   if (value->is_InlineType()) {
1297     // Null checking a scalarized but nullable inline type. Check the null marker
1298     // input instead of the oop input to avoid keeping buffer allocations alive.
1299     InlineTypeNode* vtptr = value->as_InlineType();
1300     while (vtptr->get_oop()->is_InlineType()) {
1301       vtptr = vtptr->get_oop()->as_InlineType();
1302     }
1303     null_check_common(vtptr->get_null_marker(), T_INT, assert_null, null_control, speculative, true);
1304     if (stopped()) {
1305       return top();
1306     }
1307     if (assert_null) {
1308       // TODO 8284443 Scalarize here (this currently leads to compilation bailouts)
1309       // vtptr = InlineTypeNode::make_null(_gvn, vtptr->type()->inline_klass());
1310       // replace_in_map(value, vtptr);
1311       // return vtptr;
1312       replace_in_map(value, null());
1313       return null();
1314     }
1315     bool do_replace_in_map = (null_control == nullptr || (*null_control) == top());
1316     return cast_not_null(value, do_replace_in_map);
1317   }
1318 
1319   // Construct null check
1320   Node *chk = nullptr;
1321   switch(type) {
1322     case T_LONG   : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1323     case T_INT    : chk = new CmpINode(value, _gvn.intcon(0)); break;
1324     case T_ARRAY  : // fall through
1325       type = T_OBJECT;  // simplify further tests
1326     case T_OBJECT : {
1327       const Type *t = _gvn.type( value );
1328 
1329       const TypeOopPtr* tp = t->isa_oopptr();
1330       if (tp != nullptr && !tp->is_loaded()
1331           // Only for do_null_check, not any of its siblings:
1332           && !assert_null && null_control == nullptr) {
1333         // Usually, any field access or invocation on an unloaded oop type
1334         // will simply fail to link, since the statically linked class is
1335         // likely also to be unloaded.  However, in -Xcomp mode, sometimes
1336         // the static class is loaded but the sharper oop type is not.
1337         // Rather than checking for this obscure case in lots of places,
1338         // we simply observe that a null check on an unloaded class
1339         // will always be followed by a nonsense operation, so we
1340         // can just issue the uncommon trap here.
1341         // Our access to the unloaded class will only be correct
1342         // after it has been loaded and initialized, which requires
1343         // a trip through the interpreter.
1344         ciKlass* klass = tp->unloaded_klass();
1345 #ifndef PRODUCT
1346         if (WizardMode) { tty->print("Null check of unloaded "); klass->print(); tty->cr(); }
1347 #endif
1348         uncommon_trap(Deoptimization::Reason_unloaded,
1349                       Deoptimization::Action_reinterpret,
1350                       klass, "!loaded");
1351         return top();
1352       }
1353 
1354       if (assert_null) {
1355         // See if the type is contained in NULL_PTR.
1356         // If so, then the value is already null.
1357         if (t->higher_equal(TypePtr::NULL_PTR)) {
1358           NOT_PRODUCT(explicit_null_checks_elided++);
1359           return value;           // Elided null assert quickly!
1360         }
1361       } else {
1362         // See if mixing in the null pointer changes type.
1363         // If so, then the null pointer was not allowed in the original
1364         // type.  In other words, "value" was not-null.
1365         if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1366           // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1367           NOT_PRODUCT(explicit_null_checks_elided++);
1368           return value;           // Elided null check quickly!
1369         }
1370       }
1371       chk = new CmpPNode( value, null() );
1372       break;
1373     }
1374 
1375     default:
1376       fatal("unexpected type: %s", type2name(type));
1377   }
1378   assert(chk != nullptr, "sanity check");
1379   chk = _gvn.transform(chk);
1380 
1381   BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1382   BoolNode *btst = new BoolNode( chk, btest);
1383   Node   *tst = _gvn.transform( btst );
1384 
1385   //-----------
1386   // if peephole optimizations occurred, a prior test existed.
1387   // If a prior test existed, maybe it dominates as we can avoid this test.
1388   if (tst != btst && type == T_OBJECT) {
1389     // At this point we want to scan up the CFG to see if we can
1390     // find an identical test (and so avoid this test altogether).
1391     Node *cfg = control();
1392     int depth = 0;
1393     while( depth < 16 ) {       // Limit search depth for speed
1394       if( cfg->Opcode() == Op_IfTrue &&
1395           cfg->in(0)->in(1) == tst ) {
1396         // Found prior test.  Use "cast_not_null" to construct an identical
1397         // CastPP (and hence hash to) as already exists for the prior test.
1398         // Return that casted value.
1399         if (assert_null) {
1400           replace_in_map(value, null());
1401           return null();  // do not issue the redundant test
1402         }
1403         Node *oldcontrol = control();
1404         set_control(cfg);
1405         Node *res = cast_not_null(value);
1406         set_control(oldcontrol);
1407         NOT_PRODUCT(explicit_null_checks_elided++);
1408         return res;
1409       }
1410       cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1411       if (cfg == nullptr)  break;  // Quit at region nodes
1412       depth++;
1413     }
1414   }
1415 
1416   //-----------
1417   // Branch to failure if null
1418   float ok_prob = PROB_MAX;  // a priori estimate:  nulls never happen
1419   Deoptimization::DeoptReason reason;
1420   if (assert_null) {
1421     reason = Deoptimization::reason_null_assert(speculative);
1422   } else if (type == T_OBJECT || null_marker_check) {
1423     reason = Deoptimization::reason_null_check(speculative);
1424   } else {
1425     reason = Deoptimization::Reason_div0_check;
1426   }
1427   // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1428   // ciMethodData::has_trap_at will return a conservative -1 if any
1429   // must-be-null assertion has failed.  This could cause performance
1430   // problems for a method after its first do_null_assert failure.
1431   // Consider using 'Reason_class_check' instead?
1432 
1433   // To cause an implicit null check, we set the not-null probability
1434   // to the maximum (PROB_MAX).  For an explicit check the probability
1435   // is set to a smaller value.
1436   if (null_control != nullptr || too_many_traps(reason)) {
1437     // probability is less likely
1438     ok_prob =  PROB_LIKELY_MAG(3);
1439   } else if (!assert_null &&
1440              (ImplicitNullCheckThreshold > 0) &&
1441              method() != nullptr &&
1442              (method()->method_data()->trap_count(reason)
1443               >= (uint)ImplicitNullCheckThreshold)) {
1444     ok_prob =  PROB_LIKELY_MAG(3);
1445   }
1446 
1447   if (null_control != nullptr) {
1448     IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1449     Node* null_true = _gvn.transform( new IfFalseNode(iff));
1450     set_control(      _gvn.transform( new IfTrueNode(iff)));
1451 #ifndef PRODUCT
1452     if (null_true == top()) {
1453       explicit_null_checks_elided++;
1454     }
1455 #endif
1456     (*null_control) = null_true;
1457   } else {
1458     BuildCutout unless(this, tst, ok_prob);
1459     // Check for optimizer eliding test at parse time
1460     if (stopped()) {
1461       // Failure not possible; do not bother making uncommon trap.
1462       NOT_PRODUCT(explicit_null_checks_elided++);
1463     } else if (assert_null) {
1464       uncommon_trap(reason,
1465                     Deoptimization::Action_make_not_entrant,
1466                     nullptr, "assert_null");
1467     } else {
1468       replace_in_map(value, zerocon(type));
1469       builtin_throw(reason);
1470     }
1471   }
1472 
1473   // Must throw exception, fall-thru not possible?
1474   if (stopped()) {
1475     return top();               // No result
1476   }
1477 
1478   if (assert_null) {
1479     // Cast obj to null on this path.
1480     replace_in_map(value, zerocon(type));
1481     return zerocon(type);
1482   }
1483 
1484   // Cast obj to not-null on this path, if there is no null_control.
1485   // (If there is a null_control, a non-null value may come back to haunt us.)
1486   if (type == T_OBJECT) {
1487     Node* cast = cast_not_null(value, false);
1488     if (null_control == nullptr || (*null_control) == top())
1489       replace_in_map(value, cast);
1490     value = cast;
1491   }
1492 
1493   return value;
1494 }
1495 
1496 //------------------------------cast_not_null----------------------------------
1497 // Cast obj to not-null on this path
1498 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1499   if (obj->is_InlineType()) {
1500     Node* vt = obj->isa_InlineType()->clone_if_required(&gvn(), map(), do_replace_in_map);
1501     vt->as_InlineType()->set_null_marker(_gvn);
1502     vt = _gvn.transform(vt);
1503     if (do_replace_in_map) {
1504       replace_in_map(obj, vt);
1505     }
1506     return vt;
1507   }
1508   const Type *t = _gvn.type(obj);
1509   const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1510   // Object is already not-null?
1511   if( t == t_not_null ) return obj;
1512 
1513   Node* cast = new CastPPNode(control(), obj,t_not_null);
1514   cast = _gvn.transform( cast );
1515 
1516   // Scan for instances of 'obj' in the current JVM mapping.
1517   // These instances are known to be not-null after the test.
1518   if (do_replace_in_map)
1519     replace_in_map(obj, cast);
1520 
1521   return cast;                  // Return casted value
1522 }
1523 
1524 Node* GraphKit::cast_to_non_larval(Node* obj) {
1525   const Type* obj_type = gvn().type(obj);
1526   if (obj->is_InlineType() || !obj_type->is_inlinetypeptr()) {
1527     return obj;
1528   }
1529 
1530   Node* new_obj = InlineTypeNode::make_from_oop(this, obj, obj_type->inline_klass());
1531   replace_in_map(obj, new_obj);
1532   return new_obj;
1533 }
1534 
1535 // Sometimes in intrinsics, we implicitly know an object is not null
1536 // (there's no actual null check) so we can cast it to not null. In
1537 // the course of optimizations, the input to the cast can become null.
1538 // In that case that data path will die and we need the control path
1539 // to become dead as well to keep the graph consistent. So we have to
1540 // add a check for null for which one branch can't be taken. It uses
1541 // an OpaqueNotNull node that will cause the check to be removed after loop
1542 // opts so the test goes away and the compiled code doesn't execute a
1543 // useless check.
1544 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1545   if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(value))) {
1546     return value;
1547   }
1548   Node* chk = _gvn.transform(new CmpPNode(value, null()));
1549   Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1550   Node* opaq = _gvn.transform(new OpaqueNotNullNode(C, tst));
1551   IfNode* iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1552   _gvn.set_type(iff, iff->Value(&_gvn));
1553   if (!tst->is_Con()) {
1554     record_for_igvn(iff);
1555   }
1556   Node *if_f = _gvn.transform(new IfFalseNode(iff));
1557   Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr));
1558   Node* halt = _gvn.transform(new HaltNode(if_f, frame, "unexpected null in intrinsic"));
1559   C->root()->add_req(halt);
1560   Node *if_t = _gvn.transform(new IfTrueNode(iff));
1561   set_control(if_t);
1562   return cast_not_null(value, do_replace_in_map);
1563 }
1564 
1565 
1566 //--------------------------replace_in_map-------------------------------------
1567 void GraphKit::replace_in_map(Node* old, Node* neww) {
1568   if (old == neww) {
1569     return;
1570   }
1571 
1572   map()->replace_edge(old, neww);
1573 
1574   // Note: This operation potentially replaces any edge
1575   // on the map.  This includes locals, stack, and monitors
1576   // of the current (innermost) JVM state.
1577 
1578   // don't let inconsistent types from profiling escape this
1579   // method
1580 
1581   const Type* told = _gvn.type(old);
1582   const Type* tnew = _gvn.type(neww);
1583 
1584   if (!tnew->higher_equal(told)) {
1585     return;
1586   }
1587 
1588   map()->record_replaced_node(old, neww);
1589 }
1590 
1591 
1592 //=============================================================================
1593 //--------------------------------memory---------------------------------------
1594 Node* GraphKit::memory(uint alias_idx) {
1595   MergeMemNode* mem = merged_memory();
1596   Node* p = mem->memory_at(alias_idx);
1597   assert(p != mem->empty_memory(), "empty");
1598   _gvn.set_type(p, Type::MEMORY);  // must be mapped
1599   return p;
1600 }
1601 
1602 //-----------------------------reset_memory------------------------------------
1603 Node* GraphKit::reset_memory() {
1604   Node* mem = map()->memory();
1605   // do not use this node for any more parsing!
1606   DEBUG_ONLY( map()->set_memory((Node*)nullptr) );
1607   return _gvn.transform( mem );
1608 }
1609 
1610 //------------------------------set_all_memory---------------------------------
1611 void GraphKit::set_all_memory(Node* newmem) {
1612   Node* mergemem = MergeMemNode::make(newmem);
1613   gvn().set_type_bottom(mergemem);
1614   map()->set_memory(mergemem);
1615 }
1616 
1617 //------------------------------set_all_memory_call----------------------------
1618 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1619   Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1620   set_all_memory(newmem);
1621 }
1622 
1623 //=============================================================================
1624 //
1625 // parser factory methods for MemNodes
1626 //
1627 // These are layered on top of the factory methods in LoadNode and StoreNode,
1628 // and integrate with the parser's memory state and _gvn engine.
1629 //
1630 
1631 // factory methods in "int adr_idx"
1632 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1633                           MemNode::MemOrd mo,
1634                           LoadNode::ControlDependency control_dependency,
1635                           bool require_atomic_access,
1636                           bool unaligned,
1637                           bool mismatched,
1638                           bool unsafe,
1639                           uint8_t barrier_data) {
1640   int adr_idx = C->get_alias_index(_gvn.type(adr)->isa_ptr());
1641   assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1642   const TypePtr* adr_type = nullptr; // debug-mode-only argument
1643   DEBUG_ONLY(adr_type = C->get_adr_type(adr_idx));
1644   Node* mem = memory(adr_idx);
1645   Node* ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, require_atomic_access, unaligned, mismatched, unsafe, barrier_data);
1646   ld = _gvn.transform(ld);
1647 
1648   if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1649     // Improve graph before escape analysis and boxing elimination.
1650     record_for_igvn(ld);
1651     if (ld->is_DecodeN()) {
1652       // Also record the actual load (LoadN) in case ld is DecodeN. In some
1653       // rare corner cases, ld->in(1) can be something other than LoadN (e.g.,
1654       // a Phi). Recording such cases is still perfectly sound, but may be
1655       // unnecessary and result in some minor IGVN overhead.
1656       record_for_igvn(ld->in(1));
1657     }
1658   }
1659   return ld;
1660 }
1661 
1662 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1663                                 MemNode::MemOrd mo,
1664                                 bool require_atomic_access,
1665                                 bool unaligned,
1666                                 bool mismatched,
1667                                 bool unsafe,
1668                                 int barrier_data) {
1669   int adr_idx = C->get_alias_index(_gvn.type(adr)->isa_ptr());
1670   assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1671   const TypePtr* adr_type = nullptr;
1672   DEBUG_ONLY(adr_type = C->get_adr_type(adr_idx));
1673   Node *mem = memory(adr_idx);
1674   Node* st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo, require_atomic_access);
1675   if (unaligned) {
1676     st->as_Store()->set_unaligned_access();
1677   }
1678   if (mismatched) {
1679     st->as_Store()->set_mismatched_access();
1680   }
1681   if (unsafe) {
1682     st->as_Store()->set_unsafe_access();
1683   }
1684   st->as_Store()->set_barrier_data(barrier_data);
1685   st = _gvn.transform(st);
1686   set_memory(st, adr_idx);
1687   // Back-to-back stores can only remove intermediate store with DU info
1688   // so push on worklist for optimizer.
1689   if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1690     record_for_igvn(st);
1691 
1692   return st;
1693 }
1694 
1695 Node* GraphKit::access_store_at(Node* obj,
1696                                 Node* adr,
1697                                 const TypePtr* adr_type,
1698                                 Node* val,
1699                                 const Type* val_type,
1700                                 BasicType bt,
1701                                 DecoratorSet decorators,
1702                                 bool safe_for_replace,
1703                                 const InlineTypeNode* vt) {
1704   // Transformation of a value which could be null pointer (CastPP #null)
1705   // could be delayed during Parse (for example, in adjust_map_after_if()).
1706   // Execute transformation here to avoid barrier generation in such case.
1707   if (_gvn.type(val) == TypePtr::NULL_PTR) {
1708     val = _gvn.makecon(TypePtr::NULL_PTR);
1709   }
1710 
1711   if (stopped()) {
1712     return top(); // Dead path ?
1713   }
1714 
1715   assert(val != nullptr, "not dead path");
1716   if (val->is_InlineType()) {
1717     // Store to non-flat field. Buffer the inline type and make sure
1718     // the store is re-executed if the allocation triggers deoptimization.
1719     PreserveReexecuteState preexecs(this);
1720     jvms()->set_should_reexecute(true);
1721     val = val->as_InlineType()->buffer(this, safe_for_replace);
1722   }
1723 
1724   C2AccessValuePtr addr(adr, adr_type);
1725   C2AccessValue value(val, val_type);
1726   C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr, nullptr, vt);
1727   if (access.is_raw()) {
1728     return _barrier_set->BarrierSetC2::store_at(access, value);
1729   } else {
1730     return _barrier_set->store_at(access, value);
1731   }
1732 }
1733 
1734 Node* GraphKit::access_load_at(Node* obj,   // containing obj
1735                                Node* adr,   // actual address to store val at
1736                                const TypePtr* adr_type,
1737                                const Type* val_type,
1738                                BasicType bt,
1739                                DecoratorSet decorators,
1740                                Node* ctl) {
1741   if (stopped()) {
1742     return top(); // Dead path ?
1743   }
1744 
1745   C2AccessValuePtr addr(adr, adr_type);
1746   C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr, ctl);
1747   if (access.is_raw()) {
1748     return _barrier_set->BarrierSetC2::load_at(access, val_type);
1749   } else {
1750     return _barrier_set->load_at(access, val_type);
1751   }
1752 }
1753 
1754 Node* GraphKit::access_load(Node* adr,   // actual address to load val at
1755                             const Type* val_type,
1756                             BasicType bt,
1757                             DecoratorSet decorators) {
1758   if (stopped()) {
1759     return top(); // Dead path ?
1760   }
1761 
1762   C2AccessValuePtr addr(adr, adr->bottom_type()->is_ptr());
1763   C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, nullptr, addr);
1764   if (access.is_raw()) {
1765     return _barrier_set->BarrierSetC2::load_at(access, val_type);
1766   } else {
1767     return _barrier_set->load_at(access, val_type);
1768   }
1769 }
1770 
1771 Node* GraphKit::access_atomic_cmpxchg_val_at(Node* obj,
1772                                              Node* adr,
1773                                              const TypePtr* adr_type,
1774                                              int alias_idx,
1775                                              Node* expected_val,
1776                                              Node* new_val,
1777                                              const Type* value_type,
1778                                              BasicType bt,
1779                                              DecoratorSet decorators) {
1780   C2AccessValuePtr addr(adr, adr_type);
1781   C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1782                         bt, obj, addr, alias_idx);
1783   if (access.is_raw()) {
1784     return _barrier_set->BarrierSetC2::atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1785   } else {
1786     return _barrier_set->atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1787   }
1788 }
1789 
1790 Node* GraphKit::access_atomic_cmpxchg_bool_at(Node* obj,
1791                                               Node* adr,
1792                                               const TypePtr* adr_type,
1793                                               int alias_idx,
1794                                               Node* expected_val,
1795                                               Node* new_val,
1796                                               const Type* value_type,
1797                                               BasicType bt,
1798                                               DecoratorSet decorators) {
1799   C2AccessValuePtr addr(adr, adr_type);
1800   C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1801                         bt, obj, addr, alias_idx);
1802   if (access.is_raw()) {
1803     return _barrier_set->BarrierSetC2::atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1804   } else {
1805     return _barrier_set->atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1806   }
1807 }
1808 
1809 Node* GraphKit::access_atomic_xchg_at(Node* obj,
1810                                       Node* adr,
1811                                       const TypePtr* adr_type,
1812                                       int alias_idx,
1813                                       Node* new_val,
1814                                       const Type* value_type,
1815                                       BasicType bt,
1816                                       DecoratorSet decorators) {
1817   C2AccessValuePtr addr(adr, adr_type);
1818   C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1819                         bt, obj, addr, alias_idx);
1820   if (access.is_raw()) {
1821     return _barrier_set->BarrierSetC2::atomic_xchg_at(access, new_val, value_type);
1822   } else {
1823     return _barrier_set->atomic_xchg_at(access, new_val, value_type);
1824   }
1825 }
1826 
1827 Node* GraphKit::access_atomic_add_at(Node* obj,
1828                                      Node* adr,
1829                                      const TypePtr* adr_type,
1830                                      int alias_idx,
1831                                      Node* new_val,
1832                                      const Type* value_type,
1833                                      BasicType bt,
1834                                      DecoratorSet decorators) {
1835   C2AccessValuePtr addr(adr, adr_type);
1836   C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1837   if (access.is_raw()) {
1838     return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1839   } else {
1840     return _barrier_set->atomic_add_at(access, new_val, value_type);
1841   }
1842 }
1843 
1844 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) {
1845   return _barrier_set->clone(this, src, dst, size, is_array);
1846 }
1847 
1848 //-------------------------array_element_address-------------------------
1849 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1850                                       const TypeInt* sizetype, Node* ctrl) {
1851   const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr();
1852   uint shift;
1853   uint header;
1854   if (arytype->is_flat() && arytype->klass_is_exact()) {
1855     // We can only determine the flat array layout statically if the klass is exact. Otherwise, we could have different
1856     // value classes at runtime with a potentially different layout. The caller needs to fall back to call
1857     // load/store_unknown_inline_Type() at runtime. We could return a sentinel node for the non-exact case but that
1858     // might mess with other GVN transformations in between. Thus, we just continue in the else branch normally, even
1859     // though we don't need the address node in this case and throw it away again.
1860     shift = arytype->flat_log_elem_size();
1861     header = arrayOopDesc::base_offset_in_bytes(T_FLAT_ELEMENT);
1862   } else {
1863     shift = exact_log2(type2aelembytes(elembt));
1864     header = arrayOopDesc::base_offset_in_bytes(elembt);
1865   }
1866 
1867   // short-circuit a common case (saves lots of confusing waste motion)
1868   jint idx_con = find_int_con(idx, -1);
1869   if (idx_con >= 0) {
1870     intptr_t offset = header + ((intptr_t)idx_con << shift);
1871     return basic_plus_adr(ary, offset);
1872   }
1873 
1874   // must be correct type for alignment purposes
1875   Node* base  = basic_plus_adr(ary, header);
1876   idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1877   Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1878   return basic_plus_adr(ary, base, scale);
1879 }
1880 
1881 Node* GraphKit::cast_to_flat_array(Node* array, ciInlineKlass* elem_vk) {
1882   assert(elem_vk->maybe_flat_in_array(), "no flat array for %s", elem_vk->name()->as_utf8());
1883   if (!elem_vk->has_atomic_layout() && !elem_vk->has_nullable_atomic_layout()) {
1884     return cast_to_flat_array_exact(array, elem_vk, true, false);
1885   } else if (!elem_vk->has_nullable_atomic_layout() && !elem_vk->has_non_atomic_layout()) {
1886     return cast_to_flat_array_exact(array, elem_vk, true, true);
1887   } else if (!elem_vk->has_atomic_layout() && !elem_vk->has_non_atomic_layout()) {
1888     return cast_to_flat_array_exact(array, elem_vk, false, true);
1889   }
1890 
1891   bool is_null_free = false;
1892   if (!elem_vk->has_nullable_atomic_layout()) {
1893     // Element does not have a nullable flat layout, cannot be nullable
1894     is_null_free = true;
1895   }
1896 
1897   ciArrayKlass* array_klass = ciObjArrayKlass::make(elem_vk, false);
1898   const TypeAryPtr* arytype = TypeOopPtr::make_from_klass(array_klass)->isa_aryptr();
1899   arytype = arytype->cast_to_flat(true)->cast_to_null_free(is_null_free);
1900   return _gvn.transform(new CheckCastPPNode(control(), array, arytype, ConstraintCastNode::DependencyType::NonFloatingNarrowing));
1901 }
1902 
1903 Node* GraphKit::cast_to_flat_array_exact(Node* array, ciInlineKlass* elem_vk, bool is_null_free, bool is_atomic) {
1904   assert(is_null_free || is_atomic, "nullable arrays must be atomic");
1905   ciArrayKlass* array_klass = ciObjArrayKlass::make(elem_vk, true, is_null_free, is_atomic);
1906   const TypeAryPtr* arytype = TypeOopPtr::make_from_klass(array_klass)->isa_aryptr();
1907   assert(arytype->klass_is_exact(), "inconsistency");
1908   assert(arytype->is_flat(), "inconsistency");
1909   assert(arytype->is_null_free() == is_null_free, "inconsistency");
1910   assert(arytype->is_not_null_free() == !is_null_free, "inconsistency");
1911   return _gvn.transform(new CheckCastPPNode(control(), array, arytype, ConstraintCastNode::DependencyType::NonFloatingNarrowing));
1912 }
1913 
1914 //-------------------------load_array_element-------------------------
1915 Node* GraphKit::load_array_element(Node* ary, Node* idx, const TypeAryPtr* arytype, bool set_ctrl) {
1916   const Type* elemtype = arytype->elem();
1917   BasicType elembt = elemtype->array_element_basic_type();
1918   Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1919   if (elembt == T_NARROWOOP) {
1920     elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1921   }
1922   Node* ld = access_load_at(ary, adr, arytype, elemtype, elembt,
1923                             IN_HEAP | IS_ARRAY | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0));
1924   return ld;
1925 }
1926 
1927 //-------------------------set_arguments_for_java_call-------------------------
1928 // Arguments (pre-popped from the stack) are taken from the JVMS.
1929 void GraphKit::set_arguments_for_java_call(CallJavaNode* call, bool is_late_inline) {
1930   PreserveReexecuteState preexecs(this);
1931   if (Arguments::is_valhalla_enabled()) {
1932     // Make sure the call is "re-executed", if buffering of inline type arguments triggers deoptimization.
1933     // At this point, the call hasn't been executed yet, so we will only ever execute the call once.
1934     jvms()->set_should_reexecute(true);
1935     int arg_size = method()->get_declared_signature_at_bci(bci())->arg_size_for_bc(java_bc());
1936     inc_sp(arg_size);
1937   }
1938   // Add the call arguments
1939   const TypeTuple* domain = call->tf()->domain_sig();
1940   uint nargs = domain->cnt();
1941   int arg_num = 0;
1942   for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) {
1943     Node* arg = argument(i-TypeFunc::Parms);
1944     const Type* t = domain->field_at(i);
1945     // TODO 8284443 A static call to a mismatched method should still be scalarized
1946     if (t->is_inlinetypeptr() && !call->method()->get_Method()->mismatch() && call->method()->is_scalarized_arg(arg_num)) {
1947       // We don't pass inline type arguments by reference but instead pass each field of the inline type
1948       if (!arg->is_InlineType()) {
1949         assert(_gvn.type(arg)->is_zero_type() && !t->inline_klass()->is_null_free(), "Unexpected argument type");
1950         arg = InlineTypeNode::make_from_oop(this, arg, t->inline_klass());
1951       }
1952       InlineTypeNode* vt = arg->as_InlineType();
1953       vt->pass_fields(this, call, idx, true, !t->maybe_null());
1954       // If an inline type argument is passed as fields, attach the Method* to the call site
1955       // to be able to access the extended signature later via attached_method_before_pc().
1956       // For example, see CompiledMethod::preserve_callee_argument_oops().
1957       call->set_override_symbolic_info(true);
1958       // Register an calling convention dependency on the callee method to make sure that this method is deoptimized and
1959       // re-compiled with a non-scalarized calling convention if the callee method is later marked as mismatched.
1960       C->dependencies()->assert_mismatch_calling_convention(call->method());
1961       arg_num++;
1962       continue;
1963     } else if (arg->is_InlineType()) {
1964       // Pass inline type argument via oop to callee
1965       arg = arg->as_InlineType()->buffer(this, true);
1966     }
1967     if (t != Type::HALF) {
1968       arg_num++;
1969     }
1970     call->init_req(idx++, arg);
1971   }
1972 }
1973 
1974 //---------------------------set_edges_for_java_call---------------------------
1975 // Connect a newly created call into the current JVMS.
1976 // A return value node (if any) is returned from set_edges_for_java_call.
1977 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1978 
1979   // Add the predefined inputs:
1980   call->init_req( TypeFunc::Control, control() );
1981   call->init_req( TypeFunc::I_O    , i_o() );
1982   call->init_req( TypeFunc::Memory , reset_memory() );
1983   call->init_req( TypeFunc::FramePtr, frameptr() );
1984   call->init_req( TypeFunc::ReturnAdr, top() );
1985 
1986   add_safepoint_edges(call, must_throw);
1987 
1988   Node* xcall = _gvn.transform(call);
1989 
1990   if (xcall == top()) {
1991     set_control(top());
1992     return;
1993   }
1994   assert(xcall == call, "call identity is stable");
1995 
1996   // Re-use the current map to produce the result.
1997 
1998   set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1999   set_i_o(    _gvn.transform(new ProjNode(call, TypeFunc::I_O    , separate_io_proj)));
2000   set_all_memory_call(xcall, separate_io_proj);
2001 
2002   //return xcall;   // no need, caller already has it
2003 }
2004 
2005 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
2006   if (stopped())  return top();  // maybe the call folded up?
2007 
2008   // Note:  Since any out-of-line call can produce an exception,
2009   // we always insert an I_O projection from the call into the result.
2010 
2011   make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
2012 
2013   if (separate_io_proj) {
2014     // The caller requested separate projections be used by the fall
2015     // through and exceptional paths, so replace the projections for
2016     // the fall through path.
2017     set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
2018     set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
2019   }
2020 
2021   // Capture the return value, if any.
2022   Node* ret;
2023   if (call->method() == nullptr || call->method()->return_type()->basic_type() == T_VOID) {
2024     ret = top();
2025   } else if (call->tf()->returns_inline_type_as_fields()) {
2026     // Return of multiple values (inline type fields): we create a
2027     // InlineType node, each field is a projection from the call.
2028     ciInlineKlass* vk = call->method()->return_type()->as_inline_klass();
2029     uint base_input = TypeFunc::Parms;
2030     ret = InlineTypeNode::make_from_multi(this, call, vk, base_input, false, false);
2031   } else {
2032     ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
2033     ciType* t = call->method()->return_type();
2034     if (!t->is_loaded() && InlineTypeReturnedAsFields) {
2035       // The return type is unloaded but the callee might later be C2 compiled and then return
2036       // in scalarized form when the return type is loaded. Handle this similar to what we do in
2037       // PhaseMacroExpand::expand_mh_intrinsic_return by calling into the runtime to buffer.
2038       // It's a bit unfortunate because we will deopt anyway but the interpreter needs an oop.
2039       IdealKit ideal(this);
2040       IdealVariable res(ideal);
2041       ideal.declarations_done();
2042       ideal.if_then(ret, BoolTest::eq, ideal.makecon(TypePtr::NULL_PTR)); {
2043         // Return value is null
2044         ideal.set(res, makecon(TypePtr::NULL_PTR));
2045       } ideal.else_(); {
2046         // Return value is non-null
2047         sync_kit(ideal);
2048 
2049         // Change return type of call to scalarized return
2050         const TypeFunc* tf = call->_tf;
2051         const TypeTuple* domain = OptoRuntime::store_inline_type_fields_Type()->domain_cc();
2052         const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain);
2053         call->_tf = new_tf;
2054         _gvn.set_type(call, call->Value(&_gvn));
2055         _gvn.set_type(ret, ret->Value(&_gvn));
2056 
2057         Node* store_to_buf_call = make_runtime_call(RC_NO_LEAF | RC_NO_IO,
2058                                                     OptoRuntime::store_inline_type_fields_Type(),
2059                                                     StubRoutines::store_inline_type_fields_to_buf(),
2060                                                     nullptr, TypePtr::BOTTOM, ret);
2061 
2062         // We don't know how many values are returned. This assumes the
2063         // worst case, that all available registers are used.
2064         for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2065           if (domain->field_at(i) == Type::HALF) {
2066             store_to_buf_call->init_req(i, top());
2067             continue;
2068           }
2069           Node* proj =_gvn.transform(new ProjNode(call, i));
2070           store_to_buf_call->init_req(i, proj);
2071         }
2072         make_slow_call_ex(store_to_buf_call, env()->Throwable_klass(), false);
2073 
2074         Node* buf = _gvn.transform(new ProjNode(store_to_buf_call, TypeFunc::Parms));
2075         const Type* buf_type = TypeOopPtr::make_from_klass(t->as_klass())->join_speculative(TypePtr::NOTNULL);
2076         buf = _gvn.transform(new CheckCastPPNode(control(), buf, buf_type));
2077 
2078         ideal.set(res, buf);
2079         ideal.sync_kit(this);
2080       } ideal.end_if();
2081       sync_kit(ideal);
2082       ret = _gvn.transform(ideal.value(res));
2083     }
2084     if (t->is_klass()) {
2085       const Type* type = TypeOopPtr::make_from_klass(t->as_klass());
2086       if (type->is_inlinetypeptr()) {
2087         ret = InlineTypeNode::make_from_oop(this, ret, type->inline_klass());
2088       }
2089     }
2090   }
2091 
2092   return ret;
2093 }
2094 
2095 //--------------------set_predefined_input_for_runtime_call--------------------
2096 // Reading and setting the memory state is way conservative here.
2097 // The real problem is that I am not doing real Type analysis on memory,
2098 // so I cannot distinguish card mark stores from other stores.  Across a GC
2099 // point the Store Barrier and the card mark memory has to agree.  I cannot
2100 // have a card mark store and its barrier split across the GC point from
2101 // either above or below.  Here I get that to happen by reading ALL of memory.
2102 // A better answer would be to separate out card marks from other memory.
2103 // For now, return the input memory state, so that it can be reused
2104 // after the call, if this call has restricted memory effects.
2105 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
2106   // Set fixed predefined input arguments
2107   call->init_req(TypeFunc::Control, control());
2108   call->init_req(TypeFunc::I_O, top()); // does no i/o
2109   call->init_req(TypeFunc::ReturnAdr, top());
2110   if (call->is_CallLeafPure()) {
2111     call->init_req(TypeFunc::Memory, top());
2112     call->init_req(TypeFunc::FramePtr, top());
2113     return nullptr;
2114   } else {
2115     Node* memory = reset_memory();
2116     Node* m = narrow_mem == nullptr ? memory : narrow_mem;
2117     call->init_req(TypeFunc::Memory, m); // may gc ptrs
2118     call->init_req(TypeFunc::FramePtr, frameptr());
2119     return memory;
2120   }
2121 }
2122 
2123 //-------------------set_predefined_output_for_runtime_call--------------------
2124 // Set control and memory (not i_o) from the call.
2125 // If keep_mem is not null, use it for the output state,
2126 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
2127 // If hook_mem is null, this call produces no memory effects at all.
2128 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
2129 // then only that memory slice is taken from the call.
2130 // In the last case, we must put an appropriate memory barrier before
2131 // the call, so as to create the correct anti-dependencies on loads
2132 // preceding the call.
2133 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
2134                                                       Node* keep_mem,
2135                                                       const TypePtr* hook_mem) {
2136   // no i/o
2137   set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) ));
2138   if (call->is_CallLeafPure()) {
2139     // Pure function have only control (for now) and data output, in particular
2140     // they don't touch the memory, so we don't want a memory proj that is set after.
2141     return;
2142   }
2143   if (keep_mem) {
2144     // First clone the existing memory state
2145     set_all_memory(keep_mem);
2146     if (hook_mem != nullptr) {
2147       // Make memory for the call
2148       Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) );
2149       // Set the RawPtr memory state only.  This covers all the heap top/GC stuff
2150       // We also use hook_mem to extract specific effects from arraycopy stubs.
2151       set_memory(mem, hook_mem);
2152     }
2153     // ...else the call has NO memory effects.
2154 
2155     // Make sure the call advertises its memory effects precisely.
2156     // This lets us build accurate anti-dependences in gcm.cpp.
2157     assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
2158            "call node must be constructed correctly");
2159   } else {
2160     assert(hook_mem == nullptr, "");
2161     // This is not a "slow path" call; all memory comes from the call.
2162     set_all_memory_call(call);
2163   }
2164 }
2165 
2166 // Keep track of MergeMems feeding into other MergeMems
2167 static void add_mergemem_users_to_worklist(Unique_Node_List& wl, Node* mem) {
2168   if (!mem->is_MergeMem()) {
2169     return;
2170   }
2171   for (SimpleDUIterator i(mem); i.has_next(); i.next()) {
2172     Node* use = i.get();
2173     if (use->is_MergeMem()) {
2174       wl.push(use);
2175     }
2176   }
2177 }
2178 
2179 // Replace the call with the current state of the kit.
2180 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes, bool do_asserts) {
2181   JVMState* ejvms = nullptr;
2182   if (has_exceptions()) {
2183     ejvms = transfer_exceptions_into_jvms();
2184   }
2185 
2186   ReplacedNodes replaced_nodes = map()->replaced_nodes();
2187   ReplacedNodes replaced_nodes_exception;
2188   Node* ex_ctl = top();
2189 
2190   SafePointNode* final_state = stop();
2191 
2192   // Find all the needed outputs of this call
2193   CallProjections* callprojs = call->extract_projections(true, do_asserts);
2194 
2195   Unique_Node_List wl;
2196   Node* init_mem = call->in(TypeFunc::Memory);
2197   Node* final_mem = final_state->in(TypeFunc::Memory);
2198   Node* final_ctl = final_state->in(TypeFunc::Control);
2199   Node* final_io = final_state->in(TypeFunc::I_O);
2200 
2201   // Replace all the old call edges with the edges from the inlining result
2202   if (callprojs->fallthrough_catchproj != nullptr) {
2203     C->gvn_replace_by(callprojs->fallthrough_catchproj, final_ctl);
2204   }
2205   if (callprojs->fallthrough_memproj != nullptr) {
2206     if (final_mem->is_MergeMem()) {
2207       // Parser's exits MergeMem was not transformed but may be optimized
2208       final_mem = _gvn.transform(final_mem);
2209     }
2210     C->gvn_replace_by(callprojs->fallthrough_memproj,   final_mem);
2211     add_mergemem_users_to_worklist(wl, final_mem);
2212   }
2213   if (callprojs->fallthrough_ioproj != nullptr) {
2214     C->gvn_replace_by(callprojs->fallthrough_ioproj,    final_io);
2215   }
2216 
2217   // Replace the result with the new result if it exists and is used
2218   if (callprojs->resproj[0] != nullptr && result != nullptr) {
2219     // If the inlined code is dead, the result projections for an inline type returned as
2220     // fields have not been replaced. They will go away once the call is replaced by TOP below.
2221     assert(callprojs->nb_resproj == 1 || (call->tf()->returns_inline_type_as_fields() && stopped()),
2222            "unexpected number of results");
2223     C->gvn_replace_by(callprojs->resproj[0], result);
2224   }
2225 
2226   if (ejvms == nullptr) {
2227     // No exception edges to simply kill off those paths
2228     if (callprojs->catchall_catchproj != nullptr) {
2229       C->gvn_replace_by(callprojs->catchall_catchproj, C->top());
2230     }
2231     if (callprojs->catchall_memproj != nullptr) {
2232       C->gvn_replace_by(callprojs->catchall_memproj,   C->top());
2233     }
2234     if (callprojs->catchall_ioproj != nullptr) {
2235       C->gvn_replace_by(callprojs->catchall_ioproj,    C->top());
2236     }
2237     // Replace the old exception object with top
2238     if (callprojs->exobj != nullptr) {
2239       C->gvn_replace_by(callprojs->exobj, C->top());
2240     }
2241   } else {
2242     GraphKit ekit(ejvms);
2243 
2244     // Load my combined exception state into the kit, with all phis transformed:
2245     SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
2246     replaced_nodes_exception = ex_map->replaced_nodes();
2247 
2248     Node* ex_oop = ekit.use_exception_state(ex_map);
2249 
2250     if (callprojs->catchall_catchproj != nullptr) {
2251       C->gvn_replace_by(callprojs->catchall_catchproj, ekit.control());
2252       ex_ctl = ekit.control();
2253     }
2254     if (callprojs->catchall_memproj != nullptr) {
2255       Node* ex_mem = ekit.reset_memory();
2256       C->gvn_replace_by(callprojs->catchall_memproj,   ex_mem);
2257       add_mergemem_users_to_worklist(wl, ex_mem);
2258     }
2259     if (callprojs->catchall_ioproj != nullptr) {
2260       C->gvn_replace_by(callprojs->catchall_ioproj,    ekit.i_o());
2261     }
2262 
2263     // Replace the old exception object with the newly created one
2264     if (callprojs->exobj != nullptr) {
2265       C->gvn_replace_by(callprojs->exobj, ex_oop);
2266     }
2267   }
2268 
2269   // Disconnect the call from the graph
2270   call->disconnect_inputs(C);
2271   C->gvn_replace_by(call, C->top());
2272 
2273   // Clean up any MergeMems that feed other MergeMems since the
2274   // optimizer doesn't like that.
2275   while (wl.size() > 0) {
2276     _gvn.transform(wl.pop());
2277   }
2278 
2279   if (callprojs->fallthrough_catchproj != nullptr && !final_ctl->is_top() && do_replaced_nodes) {
2280     replaced_nodes.apply(C, final_ctl);
2281   }
2282   if (!ex_ctl->is_top() && do_replaced_nodes) {
2283     replaced_nodes_exception.apply(C, ex_ctl);
2284   }
2285 }
2286 
2287 
2288 //------------------------------increment_counter------------------------------
2289 // for statistics: increment a VM counter by 1
2290 
2291 void GraphKit::increment_counter(address counter_addr) {
2292   Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2293   increment_counter(adr1);
2294 }
2295 
2296 void GraphKit::increment_counter(Node* counter_addr) {
2297   Node* ctrl = control();
2298   Node* cnt  = make_load(ctrl, counter_addr, TypeLong::LONG, T_LONG, MemNode::unordered);
2299   Node* incr = _gvn.transform(new AddLNode(cnt, _gvn.longcon(1)));
2300   store_to_memory(ctrl, counter_addr, incr, T_LONG, MemNode::unordered);
2301 }
2302 
2303 
2304 //------------------------------uncommon_trap----------------------------------
2305 // Bail out to the interpreter in mid-method.  Implemented by calling the
2306 // uncommon_trap blob.  This helper function inserts a runtime call with the
2307 // right debug info.
2308 Node* GraphKit::uncommon_trap(int trap_request,
2309                              ciKlass* klass, const char* comment,
2310                              bool must_throw,
2311                              bool keep_exact_action) {
2312   if (failing_internal()) {
2313     stop();
2314   }
2315   if (stopped())  return nullptr; // trap reachable?
2316 
2317   // Note:  If ProfileTraps is true, and if a deopt. actually
2318   // occurs here, the runtime will make sure an MDO exists.  There is
2319   // no need to call method()->ensure_method_data() at this point.
2320 
2321   // Set the stack pointer to the right value for reexecution:
2322   set_sp(reexecute_sp());
2323 
2324 #ifdef ASSERT
2325   if (!must_throw) {
2326     // Make sure the stack has at least enough depth to execute
2327     // the current bytecode.
2328     int inputs, ignored_depth;
2329     if (compute_stack_effects(inputs, ignored_depth)) {
2330       assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
2331              Bytecodes::name(java_bc()), sp(), inputs);
2332     }
2333   }
2334 #endif
2335 
2336   Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
2337   Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
2338 
2339   switch (action) {
2340   case Deoptimization::Action_maybe_recompile:
2341   case Deoptimization::Action_reinterpret:
2342     // Temporary fix for 6529811 to allow virtual calls to be sure they
2343     // get the chance to go from mono->bi->mega
2344     if (!keep_exact_action &&
2345         Deoptimization::trap_request_index(trap_request) < 0 &&
2346         too_many_recompiles(reason)) {
2347       // This BCI is causing too many recompilations.
2348       if (C->log() != nullptr) {
2349         C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
2350                 Deoptimization::trap_reason_name(reason),
2351                 Deoptimization::trap_action_name(action));
2352       }
2353       action = Deoptimization::Action_none;
2354       trap_request = Deoptimization::make_trap_request(reason, action);
2355     } else {
2356       C->set_trap_can_recompile(true);
2357     }
2358     break;
2359   case Deoptimization::Action_make_not_entrant:
2360     C->set_trap_can_recompile(true);
2361     break;
2362   case Deoptimization::Action_none:
2363   case Deoptimization::Action_make_not_compilable:
2364     break;
2365   default:
2366 #ifdef ASSERT
2367     fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action));
2368 #endif
2369     break;
2370   }
2371 
2372   if (TraceOptoParse) {
2373     char buf[100];
2374     tty->print_cr("Uncommon trap %s at bci:%d",
2375                   Deoptimization::format_trap_request(buf, sizeof(buf),
2376                                                       trap_request), bci());
2377   }
2378 
2379   CompileLog* log = C->log();
2380   if (log != nullptr) {
2381     int kid = (klass == nullptr)? -1: log->identify(klass);
2382     log->begin_elem("uncommon_trap bci='%d'", bci());
2383     char buf[100];
2384     log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2385                                                           trap_request));
2386     if (kid >= 0)         log->print(" klass='%d'", kid);
2387     if (comment != nullptr)  log->print(" comment='%s'", comment);
2388     log->end_elem();
2389   }
2390 
2391   // Make sure any guarding test views this path as very unlikely
2392   Node *i0 = control()->in(0);
2393   if (i0 != nullptr && i0->is_If()) {        // Found a guarding if test?
2394     IfNode *iff = i0->as_If();
2395     float f = iff->_prob;   // Get prob
2396     if (control()->Opcode() == Op_IfTrue) {
2397       if (f > PROB_UNLIKELY_MAG(4))
2398         iff->_prob = PROB_MIN;
2399     } else {
2400       if (f < PROB_LIKELY_MAG(4))
2401         iff->_prob = PROB_MAX;
2402     }
2403   }
2404 
2405   // Clear out dead values from the debug info.
2406   kill_dead_locals();
2407 
2408   // Now insert the uncommon trap subroutine call
2409   address call_addr = OptoRuntime::uncommon_trap_blob()->entry_point();
2410   const TypePtr* no_memory_effects = nullptr;
2411   // Pass the index of the class to be loaded
2412   Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2413                                  (must_throw ? RC_MUST_THROW : 0),
2414                                  OptoRuntime::uncommon_trap_Type(),
2415                                  call_addr, "uncommon_trap", no_memory_effects,
2416                                  intcon(trap_request));
2417   assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2418          "must extract request correctly from the graph");
2419   assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2420 
2421   call->set_req(TypeFunc::ReturnAdr, returnadr());
2422   // The debug info is the only real input to this call.
2423 
2424   // Halt-and-catch fire here.  The above call should never return!
2425   HaltNode* halt = new HaltNode(control(), frameptr(), "uncommon trap returned which should never happen"
2426                                                        PRODUCT_ONLY(COMMA /*reachable*/false));
2427   _gvn.set_type_bottom(halt);
2428   root()->add_req(halt);
2429 
2430   stop_and_kill_map();
2431   return call;
2432 }
2433 
2434 
2435 //--------------------------just_allocated_object------------------------------
2436 // Report the object that was just allocated.
2437 // It must be the case that there are no intervening safepoints.
2438 // We use this to determine if an object is so "fresh" that
2439 // it does not require card marks.
2440 Node* GraphKit::just_allocated_object(Node* current_control) {
2441   Node* ctrl = current_control;
2442   // Object::<init> is invoked after allocation, most of invoke nodes
2443   // will be reduced, but a region node is kept in parse time, we check
2444   // the pattern and skip the region node if it degraded to a copy.
2445   if (ctrl != nullptr && ctrl->is_Region() && ctrl->req() == 2 &&
2446       ctrl->as_Region()->is_copy()) {
2447     ctrl = ctrl->as_Region()->is_copy();
2448   }
2449   if (C->recent_alloc_ctl() == ctrl) {
2450    return C->recent_alloc_obj();
2451   }
2452   return nullptr;
2453 }
2454 
2455 
2456 /**
2457  * Record profiling data exact_kls for Node n with the type system so
2458  * that it can propagate it (speculation)
2459  *
2460  * @param n          node that the type applies to
2461  * @param exact_kls  type from profiling
2462  * @param maybe_null did profiling see null?
2463  *
2464  * @return           node with improved type
2465  */
2466 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2467   const Type* current_type = _gvn.type(n);
2468   assert(UseTypeSpeculation, "type speculation must be on");
2469 
2470   const TypePtr* speculative = current_type->speculative();
2471 
2472   // Should the klass from the profile be recorded in the speculative type?
2473   if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2474     const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls, Type::trust_interfaces);
2475     const TypeOopPtr* xtype = tklass->as_instance_type();
2476     assert(xtype->klass_is_exact(), "Should be exact");
2477     // Any reason to believe n is not null (from this profiling or a previous one)?
2478     assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2479     const TypePtr* ptr = (ptr_kind != ProfileNeverNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2480     // record the new speculative type's depth
2481     speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2482     speculative = speculative->with_inline_depth(jvms()->depth());
2483   } else if (current_type->would_improve_ptr(ptr_kind)) {
2484     // Profiling report that null was never seen so we can change the
2485     // speculative type to non null ptr.
2486     if (ptr_kind == ProfileAlwaysNull) {
2487       speculative = TypePtr::NULL_PTR;
2488     } else {
2489       assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2490       const TypePtr* ptr = TypePtr::NOTNULL;
2491       if (speculative != nullptr) {
2492         speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2493       } else {
2494         speculative = ptr;
2495       }
2496     }
2497   }
2498 
2499   if (speculative != current_type->speculative()) {
2500     // Build a type with a speculative type (what we think we know
2501     // about the type but will need a guard when we use it)
2502     const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative);
2503     // We're changing the type, we need a new CheckCast node to carry
2504     // the new type. The new type depends on the control: what
2505     // profiling tells us is only valid from here as far as we can
2506     // tell.
2507     Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2508     cast = _gvn.transform(cast);
2509     replace_in_map(n, cast);
2510     n = cast;
2511   }
2512 
2513   return n;
2514 }
2515 
2516 /**
2517  * Record profiling data from receiver profiling at an invoke with the
2518  * type system so that it can propagate it (speculation)
2519  *
2520  * @param n  receiver node
2521  *
2522  * @return   node with improved type
2523  */
2524 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2525   if (!UseTypeSpeculation) {
2526     return n;
2527   }
2528   ciKlass* exact_kls = profile_has_unique_klass();
2529   ProfilePtrKind ptr_kind = ProfileMaybeNull;
2530   if ((java_bc() == Bytecodes::_checkcast ||
2531        java_bc() == Bytecodes::_instanceof ||
2532        java_bc() == Bytecodes::_aastore) &&
2533       method()->method_data()->is_mature()) {
2534     ciProfileData* data = method()->method_data()->bci_to_data(bci());
2535     if (data != nullptr) {
2536       if (java_bc() == Bytecodes::_aastore) {
2537         ciKlass* array_type = nullptr;
2538         ciKlass* element_type = nullptr;
2539         ProfilePtrKind element_ptr = ProfileMaybeNull;
2540         bool flat_array = true;
2541         bool null_free_array = true;
2542         method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
2543         exact_kls = element_type;
2544         ptr_kind = element_ptr;
2545       } else {
2546         if (!data->as_BitData()->null_seen()) {
2547           ptr_kind = ProfileNeverNull;
2548         } else {
2549           if (TypeProfileCasts) {
2550             assert(data->is_ReceiverTypeData(), "bad profile data type");
2551             ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2552             uint i = 0;
2553             for (; i < call->row_limit(); i++) {
2554               ciKlass* receiver = call->receiver(i);
2555               if (receiver != nullptr) {
2556                 break;
2557               }
2558             }
2559             ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2560           }
2561         }
2562       }
2563     }
2564   }
2565   return record_profile_for_speculation(n, exact_kls, ptr_kind);
2566 }
2567 
2568 /**
2569  * Record profiling data from argument profiling at an invoke with the
2570  * type system so that it can propagate it (speculation)
2571  *
2572  * @param dest_method  target method for the call
2573  * @param bc           what invoke bytecode is this?
2574  */
2575 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2576   if (!UseTypeSpeculation) {
2577     return;
2578   }
2579   const TypeFunc* tf    = TypeFunc::make(dest_method);
2580   int             nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2581   int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2582   for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2583     const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2584     if (is_reference_type(targ->basic_type())) {
2585       ProfilePtrKind ptr_kind = ProfileMaybeNull;
2586       ciKlass* better_type = nullptr;
2587       if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2588         record_profile_for_speculation(argument(j), better_type, ptr_kind);
2589       }
2590       i++;
2591     }
2592   }
2593 }
2594 
2595 /**
2596  * Record profiling data from parameter profiling at an invoke with
2597  * the type system so that it can propagate it (speculation)
2598  */
2599 void GraphKit::record_profiled_parameters_for_speculation() {
2600   if (!UseTypeSpeculation) {
2601     return;
2602   }
2603   for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2604     if (_gvn.type(local(i))->isa_oopptr()) {
2605       ProfilePtrKind ptr_kind = ProfileMaybeNull;
2606       ciKlass* better_type = nullptr;
2607       if (method()->parameter_profiled_type(j, better_type, ptr_kind)) {
2608         record_profile_for_speculation(local(i), better_type, ptr_kind);
2609       }
2610       j++;
2611     }
2612   }
2613 }
2614 
2615 /**
2616  * Record profiling data from return value profiling at an invoke with
2617  * the type system so that it can propagate it (speculation)
2618  */
2619 void GraphKit::record_profiled_return_for_speculation() {
2620   if (!UseTypeSpeculation) {
2621     return;
2622   }
2623   ProfilePtrKind ptr_kind = ProfileMaybeNull;
2624   ciKlass* better_type = nullptr;
2625   if (method()->return_profiled_type(bci(), better_type, ptr_kind)) {
2626     // If profiling reports a single type for the return value,
2627     // feed it to the type system so it can propagate it as a
2628     // speculative type
2629     record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind);
2630   }
2631 }
2632 
2633 
2634 //=============================================================================
2635 // Generate a fast path/slow path idiom.  Graph looks like:
2636 // [foo] indicates that 'foo' is a parameter
2637 //
2638 //              [in]     null
2639 //                 \    /
2640 //                  CmpP
2641 //                  Bool ne
2642 //                   If
2643 //                  /  \
2644 //              True    False-<2>
2645 //              / |
2646 //             /  cast_not_null
2647 //           Load  |    |   ^
2648 //        [fast_test]   |   |
2649 // gvn to   opt_test    |   |
2650 //          /    \      |  <1>
2651 //      True     False  |
2652 //        |         \\  |
2653 //   [slow_call]     \[fast_result]
2654 //    Ctl   Val       \      \
2655 //     |               \      \
2656 //    Catch       <1>   \      \
2657 //   /    \        ^     \      \
2658 //  Ex    No_Ex    |      \      \
2659 //  |       \   \  |       \ <2>  \
2660 //  ...      \  [slow_res] |  |    \   [null_result]
2661 //            \         \--+--+---  |  |
2662 //             \           | /    \ | /
2663 //              --------Region     Phi
2664 //
2665 //=============================================================================
2666 // Code is structured as a series of driver functions all called 'do_XXX' that
2667 // call a set of helper functions.  Helper functions first, then drivers.
2668 
2669 //------------------------------null_check_oop---------------------------------
2670 // Null check oop.  Set null-path control into Region in slot 3.
2671 // Make a cast-not-nullness use the other not-null control.  Return cast.
2672 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2673                                bool never_see_null,
2674                                bool safe_for_replace,
2675                                bool speculative) {
2676   // Initial null check taken path
2677   (*null_control) = top();
2678   Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative);
2679 
2680   // Generate uncommon_trap:
2681   if (never_see_null && (*null_control) != top()) {
2682     // If we see an unexpected null at a check-cast we record it and force a
2683     // recompile; the offending check-cast will be compiled to handle nulls.
2684     // If we see more than one offending BCI, then all checkcasts in the
2685     // method will be compiled to handle nulls.
2686     PreserveJVMState pjvms(this);
2687     set_control(*null_control);
2688     replace_in_map(value, null());
2689     Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative);
2690     uncommon_trap(reason,
2691                   Deoptimization::Action_make_not_entrant);
2692     (*null_control) = top();    // null path is dead
2693   }
2694   if ((*null_control) == top() && safe_for_replace) {
2695     replace_in_map(value, cast);
2696   }
2697 
2698   // Cast away null-ness on the result
2699   return cast;
2700 }
2701 
2702 //------------------------------opt_iff----------------------------------------
2703 // Optimize the fast-check IfNode.  Set the fast-path region slot 2.
2704 // Return slow-path control.
2705 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2706   IfNode *opt_iff = _gvn.transform(iff)->as_If();
2707 
2708   // Fast path taken; set region slot 2
2709   Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) );
2710   region->init_req(2,fast_taken); // Capture fast-control
2711 
2712   // Fast path not-taken, i.e. slow path
2713   Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) );
2714   return slow_taken;
2715 }
2716 
2717 //-----------------------------make_runtime_call-------------------------------
2718 Node* GraphKit::make_runtime_call(int flags,
2719                                   const TypeFunc* call_type, address call_addr,
2720                                   const char* call_name,
2721                                   const TypePtr* adr_type,
2722                                   // The following parms are all optional.
2723                                   // The first null ends the list.
2724                                   Node* parm0, Node* parm1,
2725                                   Node* parm2, Node* parm3,
2726                                   Node* parm4, Node* parm5,
2727                                   Node* parm6, Node* parm7) {
2728   assert(call_addr != nullptr, "must not call null targets");
2729 
2730   // Slow-path call
2731   bool is_leaf = !(flags & RC_NO_LEAF);
2732   bool has_io  = (!is_leaf && !(flags & RC_NO_IO));
2733   if (call_name == nullptr) {
2734     assert(!is_leaf, "must supply name for leaf");
2735     call_name = OptoRuntime::stub_name(call_addr);
2736   }
2737   CallNode* call;
2738   if (!is_leaf) {
2739     call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type);
2740   } else if (flags & RC_NO_FP) {
2741     call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2742   } else  if (flags & RC_VECTOR){
2743     uint num_bits = call_type->range_sig()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte;
2744     call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits);
2745   } else if (flags & RC_PURE) {
2746     assert(adr_type == nullptr, "pure call does not touch memory");
2747     call = new CallLeafPureNode(call_type, call_addr, call_name);
2748   } else {
2749     call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2750   }
2751 
2752   // The following is similar to set_edges_for_java_call,
2753   // except that the memory effects of the call are restricted to AliasIdxRaw.
2754 
2755   // Slow path call has no side-effects, uses few values
2756   bool wide_in  = !(flags & RC_NARROW_MEM);
2757   bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2758 
2759   Node* prev_mem = nullptr;
2760   if (wide_in) {
2761     prev_mem = set_predefined_input_for_runtime_call(call);
2762   } else {
2763     assert(!wide_out, "narrow in => narrow out");
2764     Node* narrow_mem = memory(adr_type);
2765     prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2766   }
2767 
2768   // Hook each parm in order.  Stop looking at the first null.
2769   if (parm0 != nullptr) { call->init_req(TypeFunc::Parms+0, parm0);
2770   if (parm1 != nullptr) { call->init_req(TypeFunc::Parms+1, parm1);
2771   if (parm2 != nullptr) { call->init_req(TypeFunc::Parms+2, parm2);
2772   if (parm3 != nullptr) { call->init_req(TypeFunc::Parms+3, parm3);
2773   if (parm4 != nullptr) { call->init_req(TypeFunc::Parms+4, parm4);
2774   if (parm5 != nullptr) { call->init_req(TypeFunc::Parms+5, parm5);
2775   if (parm6 != nullptr) { call->init_req(TypeFunc::Parms+6, parm6);
2776   if (parm7 != nullptr) { call->init_req(TypeFunc::Parms+7, parm7);
2777   /* close each nested if ===> */  } } } } } } } }
2778   assert(call->in(call->req()-1) != nullptr || (call->req()-1) > (TypeFunc::Parms+7), "must initialize all parms");
2779 
2780   if (!is_leaf) {
2781     // Non-leaves can block and take safepoints:
2782     add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2783   }
2784   // Non-leaves can throw exceptions:
2785   if (has_io) {
2786     call->set_req(TypeFunc::I_O, i_o());
2787   }
2788 
2789   if (flags & RC_UNCOMMON) {
2790     // Set the count to a tiny probability.  Cf. Estimate_Block_Frequency.
2791     // (An "if" probability corresponds roughly to an unconditional count.
2792     // Sort of.)
2793     call->set_cnt(PROB_UNLIKELY_MAG(4));
2794   }
2795 
2796   Node* c = _gvn.transform(call);
2797   assert(c == call, "cannot disappear");
2798 
2799   if (wide_out) {
2800     // Slow path call has full side-effects.
2801     set_predefined_output_for_runtime_call(call);
2802   } else {
2803     // Slow path call has few side-effects, and/or sets few values.
2804     set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2805   }
2806 
2807   if (has_io) {
2808     set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2809   }
2810   return call;
2811 
2812 }
2813 
2814 // i2b
2815 Node* GraphKit::sign_extend_byte(Node* in) {
2816   Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24)));
2817   return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24)));
2818 }
2819 
2820 // i2s
2821 Node* GraphKit::sign_extend_short(Node* in) {
2822   Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16)));
2823   return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16)));
2824 }
2825 
2826 
2827 //------------------------------merge_memory-----------------------------------
2828 // Merge memory from one path into the current memory state.
2829 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2830   for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2831     Node* old_slice = mms.force_memory();
2832     Node* new_slice = mms.memory2();
2833     if (old_slice != new_slice) {
2834       PhiNode* phi;
2835       if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2836         if (mms.is_empty()) {
2837           // clone base memory Phi's inputs for this memory slice
2838           assert(old_slice == mms.base_memory(), "sanity");
2839           phi = PhiNode::make(region, nullptr, Type::MEMORY, mms.adr_type(C));
2840           _gvn.set_type(phi, Type::MEMORY);
2841           for (uint i = 1; i < phi->req(); i++) {
2842             phi->init_req(i, old_slice->in(i));
2843           }
2844         } else {
2845           phi = old_slice->as_Phi(); // Phi was generated already
2846         }
2847       } else {
2848         phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2849         _gvn.set_type(phi, Type::MEMORY);
2850       }
2851       phi->set_req(new_path, new_slice);
2852       mms.set_memory(phi);
2853     }
2854   }
2855 }
2856 
2857 //------------------------------make_slow_call_ex------------------------------
2858 // Make the exception handler hookups for the slow call
2859 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2860   if (stopped())  return;
2861 
2862   // Make a catch node with just two handlers:  fall-through and catch-all
2863   Node* i_o  = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2864   Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) );
2865   Node* norm = new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci);
2866   _gvn.set_type_bottom(norm);
2867   C->record_for_igvn(norm);
2868   Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index,    CatchProjNode::no_handler_bci) );
2869 
2870   { PreserveJVMState pjvms(this);
2871     set_control(excp);
2872     set_i_o(i_o);
2873 
2874     if (excp != top()) {
2875       if (deoptimize) {
2876         // Deoptimize if an exception is caught. Don't construct exception state in this case.
2877         uncommon_trap(Deoptimization::Reason_unhandled,
2878                       Deoptimization::Action_none);
2879       } else {
2880         // Create an exception state also.
2881         // Use an exact type if the caller has a specific exception.
2882         const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2883         Node*       ex_oop  = new CreateExNode(ex_type, control(), i_o);
2884         add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2885       }
2886     }
2887   }
2888 
2889   // Get the no-exception control from the CatchNode.
2890   set_control(norm);
2891 }
2892 
2893 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN& gvn, BasicType bt) {
2894   Node* cmp = nullptr;
2895   switch(bt) {
2896   case T_INT: cmp = new CmpINode(in1, in2); break;
2897   case T_ADDRESS: cmp = new CmpPNode(in1, in2); break;
2898   default: fatal("unexpected comparison type %s", type2name(bt));
2899   }
2900   cmp = gvn.transform(cmp);
2901   Node* bol = gvn.transform(new BoolNode(cmp, test));
2902   IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN);
2903   gvn.transform(iff);
2904   if (!bol->is_Con()) gvn.record_for_igvn(iff);
2905   return iff;
2906 }
2907 
2908 //-------------------------------gen_subtype_check-----------------------------
2909 // Generate a subtyping check.  Takes as input the subtype and supertype.
2910 // Returns 2 values: sets the default control() to the true path and returns
2911 // the false path.  Only reads invariant memory; sets no (visible) memory.
2912 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2913 // but that's not exposed to the optimizer.  This call also doesn't take in an
2914 // Object; if you wish to check an Object you need to load the Object's class
2915 // prior to coming here.
2916 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, Node* mem, PhaseGVN& gvn,
2917                                ciMethod* method, int bci) {
2918   Compile* C = gvn.C;
2919   if ((*ctrl)->is_top()) {
2920     return C->top();
2921   }
2922 
2923   const TypeKlassPtr* klass_ptr_type = gvn.type(superklass)->is_klassptr();
2924   // For a direct pointer comparison, we need the refined array klass pointer
2925   Node* vm_superklass = superklass;
2926   if (klass_ptr_type->isa_aryklassptr() && klass_ptr_type->klass_is_exact()) {
2927     assert(!klass_ptr_type->is_aryklassptr()->is_refined_type(), "Unexpected refined array klass pointer");
2928     vm_superklass = gvn.makecon(klass_ptr_type->is_aryklassptr()->cast_to_refined_array_klass_ptr());
2929   }
2930 
2931   // Fast check for identical types, perhaps identical constants.
2932   // The types can even be identical non-constants, in cases
2933   // involving Array.newInstance, Object.clone, etc.
2934   if (subklass == superklass)
2935     return C->top();             // false path is dead; no test needed.
2936 
2937   if (gvn.type(superklass)->singleton()) {
2938     const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
2939     const TypeKlassPtr* subk   = gvn.type(subklass)->is_klassptr();
2940 
2941     // In the common case of an exact superklass, try to fold up the
2942     // test before generating code.  You may ask, why not just generate
2943     // the code and then let it fold up?  The answer is that the generated
2944     // code will necessarily include null checks, which do not always
2945     // completely fold away.  If they are also needless, then they turn
2946     // into a performance loss.  Example:
2947     //    Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2948     // Here, the type of 'fa' is often exact, so the store check
2949     // of fa[1]=x will fold up, without testing the nullness of x.
2950     //
2951     // At macro expansion, we would have already folded the SubTypeCheckNode
2952     // being expanded here because we always perform the static sub type
2953     // check in SubTypeCheckNode::sub() regardless of whether
2954     // StressReflectiveCode is set or not. We can therefore skip this
2955     // static check when StressReflectiveCode is on.
2956     switch (C->static_subtype_check(superk, subk)) {
2957     case Compile::SSC_always_false:
2958       {
2959         Node* always_fail = *ctrl;
2960         *ctrl = gvn.C->top();
2961         return always_fail;
2962       }
2963     case Compile::SSC_always_true:
2964       return C->top();
2965     case Compile::SSC_easy_test:
2966       {
2967         // Just do a direct pointer compare and be done.
2968         IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, vm_superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2969         *ctrl = gvn.transform(new IfTrueNode(iff));
2970         return gvn.transform(new IfFalseNode(iff));
2971       }
2972     case Compile::SSC_full_test:
2973       break;
2974     default:
2975       ShouldNotReachHere();
2976     }
2977   }
2978 
2979   // %%% Possible further optimization:  Even if the superklass is not exact,
2980   // if the subklass is the unique subtype of the superklass, the check
2981   // will always succeed.  We could leave a dependency behind to ensure this.
2982 
2983   // First load the super-klass's check-offset
2984   Node *p1 = gvn.transform(new AddPNode(superklass, superklass, gvn.MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2985   Node* m = C->immutable_memory();
2986   Node *chk_off = gvn.transform(new LoadINode(nullptr, m, p1, gvn.type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2987   int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2988   const TypeInt* chk_off_t = chk_off->Value(&gvn)->isa_int();
2989   int chk_off_con = (chk_off_t != nullptr && chk_off_t->is_con()) ? chk_off_t->get_con() : cacheoff_con;
2990   bool might_be_cache = (chk_off_con == cacheoff_con);
2991 
2992   // Load from the sub-klass's super-class display list, or a 1-word cache of
2993   // the secondary superclass list, or a failing value with a sentinel offset
2994   // if the super-klass is an interface or exceptionally deep in the Java
2995   // hierarchy and we have to scan the secondary superclass list the hard way.
2996   // Worst-case type is a little odd: null is allowed as a result (usually
2997   // klass loads can never produce a null).
2998   Node *chk_off_X = chk_off;
2999 #ifdef _LP64
3000   chk_off_X = gvn.transform(new ConvI2LNode(chk_off_X));
3001 #endif
3002   Node *p2 = gvn.transform(new AddPNode(subklass,subklass,chk_off_X));
3003   // For some types like interfaces the following loadKlass is from a 1-word
3004   // cache which is mutable so can't use immutable memory.  Other
3005   // types load from the super-class display table which is immutable.
3006   Node *kmem = C->immutable_memory();
3007   // secondary_super_cache is not immutable but can be treated as such because:
3008   // - no ideal node writes to it in a way that could cause an
3009   //   incorrect/missed optimization of the following Load.
3010   // - it's a cache so, worse case, not reading the latest value
3011   //   wouldn't cause incorrect execution
3012   if (might_be_cache && mem != nullptr) {
3013     kmem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(C->get_alias_index(gvn.type(p2)->is_ptr())) : mem;
3014   }
3015   Node* nkls = gvn.transform(LoadKlassNode::make(gvn, kmem, p2, gvn.type(p2)->is_ptr(), TypeInstKlassPtr::OBJECT_OR_NULL));
3016 
3017   // Compile speed common case: ARE a subtype and we canNOT fail
3018   if (superklass == nkls) {
3019     return C->top();             // false path is dead; no test needed.
3020   }
3021 
3022   // Gather the various success & failures here
3023   RegionNode* r_not_subtype = new RegionNode(3);
3024   gvn.record_for_igvn(r_not_subtype);
3025   RegionNode* r_ok_subtype = new RegionNode(4);
3026   gvn.record_for_igvn(r_ok_subtype);
3027 
3028   // If we might perform an expensive check, first try to take advantage of profile data that was attached to the
3029   // SubTypeCheck node
3030   if (might_be_cache && method != nullptr && VM_Version::profile_all_receivers_at_type_check()) {
3031     ciCallProfile profile = method->call_profile_at_bci(bci);
3032     float total_prob = 0;
3033     for (int i = 0; profile.has_receiver(i); ++i) {
3034       float prob = profile.receiver_prob(i);
3035       total_prob += prob;
3036     }
3037     if (total_prob * 100. >= TypeProfileSubTypeCheckCommonThreshold) {
3038       const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
3039       for (int i = 0; profile.has_receiver(i); ++i) {
3040         ciKlass* klass = profile.receiver(i);
3041         const TypeKlassPtr* klass_t = TypeKlassPtr::make(klass);
3042         Compile::SubTypeCheckResult result = C->static_subtype_check(superk, klass_t);
3043         if (result != Compile::SSC_always_true && result != Compile::SSC_always_false) {
3044           continue;
3045         }
3046         if (klass_t->isa_aryklassptr()) {
3047           // For a direct pointer comparison, we need the refined array klass pointer
3048           klass_t = klass_t->is_aryklassptr()->cast_to_refined_array_klass_ptr();
3049         }
3050         float prob = profile.receiver_prob(i);
3051         ConNode* klass_node = gvn.makecon(klass_t);
3052         IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, klass_node, BoolTest::eq, prob, gvn, T_ADDRESS);
3053         Node* iftrue = gvn.transform(new IfTrueNode(iff));
3054 
3055         if (result == Compile::SSC_always_true) {
3056           r_ok_subtype->add_req(iftrue);
3057         } else {
3058           assert(result == Compile::SSC_always_false, "");
3059           r_not_subtype->add_req(iftrue);
3060         }
3061         *ctrl = gvn.transform(new IfFalseNode(iff));
3062       }
3063     }
3064   }
3065 
3066   // See if we get an immediate positive hit.  Happens roughly 83% of the
3067   // time.  Test to see if the value loaded just previously from the subklass
3068   // is exactly the superklass.
3069   IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
3070   Node *iftrue1 = gvn.transform( new IfTrueNode (iff1));
3071   *ctrl = gvn.transform(new IfFalseNode(iff1));
3072 
3073   // Compile speed common case: Check for being deterministic right now.  If
3074   // chk_off is a constant and not equal to cacheoff then we are NOT a
3075   // subklass.  In this case we need exactly the 1 test above and we can
3076   // return those results immediately.
3077   if (!might_be_cache) {
3078     Node* not_subtype_ctrl = *ctrl;
3079     *ctrl = iftrue1; // We need exactly the 1 test above
3080     PhaseIterGVN* igvn = gvn.is_IterGVN();
3081     if (igvn != nullptr) {
3082       igvn->remove_globally_dead_node(r_ok_subtype);
3083       igvn->remove_globally_dead_node(r_not_subtype);
3084     }
3085     return not_subtype_ctrl;
3086   }
3087 
3088   r_ok_subtype->init_req(1, iftrue1);
3089 
3090   // Check for immediate negative hit.  Happens roughly 11% of the time (which
3091   // is roughly 63% of the remaining cases).  Test to see if the loaded
3092   // check-offset points into the subklass display list or the 1-element
3093   // cache.  If it points to the display (and NOT the cache) and the display
3094   // missed then it's not a subtype.
3095   Node *cacheoff = gvn.intcon(cacheoff_con);
3096   IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
3097   r_not_subtype->init_req(1, gvn.transform(new IfTrueNode (iff2)));
3098   *ctrl = gvn.transform(new IfFalseNode(iff2));
3099 
3100   // Check for self.  Very rare to get here, but it is taken 1/3 the time.
3101   // No performance impact (too rare) but allows sharing of secondary arrays
3102   // which has some footprint reduction.
3103   IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, vm_superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
3104   r_ok_subtype->init_req(2, gvn.transform(new IfTrueNode(iff3)));
3105   *ctrl = gvn.transform(new IfFalseNode(iff3));
3106 
3107   // -- Roads not taken here: --
3108   // We could also have chosen to perform the self-check at the beginning
3109   // of this code sequence, as the assembler does.  This would not pay off
3110   // the same way, since the optimizer, unlike the assembler, can perform
3111   // static type analysis to fold away many successful self-checks.
3112   // Non-foldable self checks work better here in second position, because
3113   // the initial primary superclass check subsumes a self-check for most
3114   // types.  An exception would be a secondary type like array-of-interface,
3115   // which does not appear in its own primary supertype display.
3116   // Finally, we could have chosen to move the self-check into the
3117   // PartialSubtypeCheckNode, and from there out-of-line in a platform
3118   // dependent manner.  But it is worthwhile to have the check here,
3119   // where it can be perhaps be optimized.  The cost in code space is
3120   // small (register compare, branch).
3121 
3122   // Now do a linear scan of the secondary super-klass array.  Again, no real
3123   // performance impact (too rare) but it's gotta be done.
3124   // Since the code is rarely used, there is no penalty for moving it
3125   // out of line, and it can only improve I-cache density.
3126   // The decision to inline or out-of-line this final check is platform
3127   // dependent, and is found in the AD file definition of PartialSubtypeCheck.
3128   Node* psc = gvn.transform(
3129     new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
3130 
3131   IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
3132   r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4)));
3133   r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4)));
3134 
3135   // Return false path; set default control to true path.
3136   *ctrl = gvn.transform(r_ok_subtype);
3137   return gvn.transform(r_not_subtype);
3138 }
3139 
3140 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) {
3141   const Type* sub_t = _gvn.type(obj_or_subklass);
3142   if (sub_t->make_oopptr() != nullptr && sub_t->make_oopptr()->is_inlinetypeptr()) {
3143     sub_t = TypeKlassPtr::make(sub_t->inline_klass());
3144     obj_or_subklass = makecon(sub_t);
3145   }
3146   bool expand_subtype_check = C->post_loop_opts_phase(); // macro node expansion is over
3147   if (expand_subtype_check) {
3148     MergeMemNode* mem = merged_memory();
3149     Node* ctrl = control();
3150     Node* subklass = obj_or_subklass;
3151     if (!sub_t->isa_klassptr()) {
3152       subklass = load_object_klass(obj_or_subklass);
3153     }
3154 
3155     Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn, method(), bci());
3156     set_control(ctrl);
3157     return n;
3158   }
3159 
3160   Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass, method(), bci()));
3161   Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq));
3162   IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3163   set_control(_gvn.transform(new IfTrueNode(iff)));
3164   return _gvn.transform(new IfFalseNode(iff));
3165 }
3166 
3167 // Profile-driven exact type check:
3168 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
3169                                     float prob, Node* *casted_receiver) {
3170   assert(!klass->is_interface(), "no exact type check on interfaces");
3171   Node* fail = top();
3172   const Type* rec_t = _gvn.type(receiver);
3173   if (rec_t->is_inlinetypeptr()) {
3174     if (klass->equals(rec_t->inline_klass())) {
3175       (*casted_receiver) = receiver; // Always passes
3176     } else {
3177       (*casted_receiver) = top();    // Always fails
3178       fail = control();
3179       set_control(top());
3180     }
3181     return fail;
3182   }
3183   const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces);
3184   if (tklass->isa_aryklassptr()) {
3185     // For a direct pointer comparison, we need the refined array klass pointer
3186     tklass = tklass->is_aryklassptr()->cast_to_refined_array_klass_ptr();
3187   }
3188   Node* recv_klass = load_object_klass(receiver);
3189   fail = type_check(recv_klass, tklass, prob);
3190 
3191   if (!stopped()) {
3192     const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3193     const TypeOopPtr* recv_xtype = tklass->as_instance_type();
3194     assert(recv_xtype->klass_is_exact(), "");
3195 
3196     if (!receiver_type->higher_equal(recv_xtype)) { // ignore redundant casts
3197       // Subsume downstream occurrences of receiver with a cast to
3198       // recv_xtype, since now we know what the type will be.
3199       Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
3200       Node* res = _gvn.transform(cast);
3201       if (recv_xtype->is_inlinetypeptr()) {
3202         assert(!gvn().type(res)->maybe_null(), "receiver should never be null");
3203         res = InlineTypeNode::make_from_oop(this, res, recv_xtype->inline_klass());
3204       }
3205       (*casted_receiver) = res;
3206       assert(!(*casted_receiver)->is_top(), "that path should be unreachable");
3207       // (User must make the replace_in_map call.)
3208     }
3209   }
3210 
3211   return fail;
3212 }
3213 
3214 Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass,
3215                            float prob) {
3216   Node* want_klass = makecon(tklass);
3217   Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass));
3218   Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3219   IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
3220   set_control(_gvn.transform(new IfTrueNode (iff)));
3221   Node* fail = _gvn.transform(new IfFalseNode(iff));
3222   return fail;
3223 }
3224 
3225 //------------------------------subtype_check_receiver-------------------------
3226 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
3227                                        Node** casted_receiver) {
3228   const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces)->try_improve();
3229   Node* want_klass = makecon(tklass);
3230 
3231   Node* slow_ctl = gen_subtype_check(receiver, want_klass);
3232 
3233   // Ignore interface type information until interface types are properly tracked.
3234   if (!stopped() && !klass->is_interface()) {
3235     const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3236     const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
3237     if (receiver_type != nullptr && !receiver_type->higher_equal(recv_type)) { // ignore redundant casts
3238       Node* cast = _gvn.transform(new CheckCastPPNode(control(), receiver, recv_type));
3239       if (recv_type->is_inlinetypeptr()) {
3240         cast = InlineTypeNode::make_from_oop(this, cast, recv_type->inline_klass());
3241       }
3242       (*casted_receiver) = cast;
3243     }
3244   }
3245 
3246   return slow_ctl;
3247 }
3248 
3249 //------------------------------seems_never_null-------------------------------
3250 // Use null_seen information if it is available from the profile.
3251 // If we see an unexpected null at a type check we record it and force a
3252 // recompile; the offending check will be recompiled to handle nulls.
3253 // If we see several offending BCIs, then all checks in the
3254 // method will be recompiled.
3255 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
3256   speculating = !_gvn.type(obj)->speculative_maybe_null();
3257   Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
3258   if (UncommonNullCast               // Cutout for this technique
3259       && obj != null()               // And not the -Xcomp stupid case?
3260       && !too_many_traps(reason)
3261       ) {
3262     if (speculating) {
3263       return true;
3264     }
3265     if (data == nullptr)
3266       // Edge case:  no mature data.  Be optimistic here.
3267       return true;
3268     // If the profile has not seen a null, assume it won't happen.
3269     assert(java_bc() == Bytecodes::_checkcast ||
3270            java_bc() == Bytecodes::_instanceof ||
3271            java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
3272     return !data->as_BitData()->null_seen();
3273   }
3274   speculating = false;
3275   return false;
3276 }
3277 
3278 void GraphKit::guard_klass_being_initialized(Node* klass) {
3279   int init_state_off = in_bytes(InstanceKlass::init_state_offset());
3280   Node* adr = basic_plus_adr(top(), klass, init_state_off);
3281   Node* init_state = LoadNode::make(_gvn, nullptr, immutable_memory(), adr,
3282                                     adr->bottom_type()->is_ptr(), TypeInt::BYTE,
3283                                     T_BYTE, MemNode::acquire);
3284   init_state = _gvn.transform(init_state);
3285 
3286   Node* being_initialized_state = makecon(TypeInt::make(InstanceKlass::being_initialized));
3287 
3288   Node* chk = _gvn.transform(new CmpINode(being_initialized_state, init_state));
3289   Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
3290 
3291   { BuildCutout unless(this, tst, PROB_MAX);
3292     uncommon_trap(Deoptimization::Reason_initialized, Deoptimization::Action_reinterpret);
3293   }
3294 }
3295 
3296 void GraphKit::guard_init_thread(Node* klass) {
3297   int init_thread_off = in_bytes(InstanceKlass::init_thread_offset());
3298   Node* adr = basic_plus_adr(top(), klass, init_thread_off);
3299 
3300   Node* init_thread = LoadNode::make(_gvn, nullptr, immutable_memory(), adr,
3301                                      adr->bottom_type()->is_ptr(), TypePtr::NOTNULL,
3302                                      T_ADDRESS, MemNode::unordered);
3303   init_thread = _gvn.transform(init_thread);
3304 
3305   Node* cur_thread = _gvn.transform(new ThreadLocalNode());
3306 
3307   Node* chk = _gvn.transform(new CmpPNode(cur_thread, init_thread));
3308   Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
3309 
3310   { BuildCutout unless(this, tst, PROB_MAX);
3311     uncommon_trap(Deoptimization::Reason_uninitialized, Deoptimization::Action_none);
3312   }
3313 }
3314 
3315 void GraphKit::clinit_barrier(ciInstanceKlass* ik, ciMethod* context) {
3316   if (ik->is_being_initialized()) {
3317     if (C->needs_clinit_barrier(ik, context)) {
3318       Node* klass = makecon(TypeKlassPtr::make(ik));
3319       guard_klass_being_initialized(klass);
3320       guard_init_thread(klass);
3321       insert_mem_bar(Op_MemBarCPUOrder);
3322     }
3323   } else if (ik->is_initialized()) {
3324     return; // no barrier needed
3325   } else {
3326     uncommon_trap(Deoptimization::Reason_uninitialized,
3327                   Deoptimization::Action_reinterpret,
3328                   nullptr);
3329   }
3330 }
3331 
3332 //------------------------maybe_cast_profiled_receiver-------------------------
3333 // If the profile has seen exactly one type, narrow to exactly that type.
3334 // Subsequent type checks will always fold up.
3335 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
3336                                              const TypeKlassPtr* require_klass,
3337                                              ciKlass* spec_klass,
3338                                              bool safe_for_replace) {
3339   if (!UseTypeProfile || !TypeProfileCasts) return nullptr;
3340 
3341   Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != nullptr);
3342 
3343   // Make sure we haven't already deoptimized from this tactic.
3344   if (too_many_traps_or_recompiles(reason))
3345     return nullptr;
3346 
3347   // (No, this isn't a call, but it's enough like a virtual call
3348   // to use the same ciMethod accessor to get the profile info...)
3349   // If we have a speculative type use it instead of profiling (which
3350   // may not help us)
3351   ciKlass* exact_kls = spec_klass;
3352   if (exact_kls == nullptr) {
3353     if (java_bc() == Bytecodes::_aastore) {
3354       ciKlass* array_type = nullptr;
3355       ciKlass* element_type = nullptr;
3356       ProfilePtrKind element_ptr = ProfileMaybeNull;
3357       bool flat_array = true;
3358       bool null_free_array = true;
3359       method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
3360       exact_kls = element_type;
3361     } else {
3362       exact_kls = profile_has_unique_klass();
3363     }
3364   }
3365   if (exact_kls != nullptr) {// no cast failures here
3366     if (require_klass == nullptr ||
3367         C->static_subtype_check(require_klass, TypeKlassPtr::make(exact_kls, Type::trust_interfaces)) == Compile::SSC_always_true) {
3368       // If we narrow the type to match what the type profile sees or
3369       // the speculative type, we can then remove the rest of the
3370       // cast.
3371       // This is a win, even if the exact_kls is very specific,
3372       // because downstream operations, such as method calls,
3373       // will often benefit from the sharper type.
3374       Node* exact_obj = not_null_obj; // will get updated in place...
3375       Node* slow_ctl  = type_check_receiver(exact_obj, exact_kls, 1.0,
3376                                             &exact_obj);
3377       { PreserveJVMState pjvms(this);
3378         set_control(slow_ctl);
3379         uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
3380       }
3381       if (safe_for_replace) {
3382         replace_in_map(not_null_obj, exact_obj);
3383       }
3384       return exact_obj;
3385     }
3386     // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us.
3387   }
3388 
3389   return nullptr;
3390 }
3391 
3392 /**
3393  * Cast obj to type and emit guard unless we had too many traps here
3394  * already
3395  *
3396  * @param obj       node being casted
3397  * @param type      type to cast the node to
3398  * @param not_null  true if we know node cannot be null
3399  */
3400 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
3401                                         ciKlass* type,
3402                                         bool not_null) {
3403   if (stopped()) {
3404     return obj;
3405   }
3406 
3407   // type is null if profiling tells us this object is always null
3408   if (type != nullptr) {
3409     Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
3410     Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check;
3411 
3412     if (!too_many_traps_or_recompiles(null_reason) &&
3413         !too_many_traps_or_recompiles(class_reason)) {
3414       Node* not_null_obj = nullptr;
3415       // not_null is true if we know the object is not null and
3416       // there's no need for a null check
3417       if (!not_null) {
3418         Node* null_ctl = top();
3419         not_null_obj = null_check_oop(obj, &null_ctl, true, true, true);
3420         assert(null_ctl->is_top(), "no null control here");
3421       } else {
3422         not_null_obj = obj;
3423       }
3424 
3425       Node* exact_obj = not_null_obj;
3426       ciKlass* exact_kls = type;
3427       Node* slow_ctl  = type_check_receiver(exact_obj, exact_kls, 1.0,
3428                                             &exact_obj);
3429       {
3430         PreserveJVMState pjvms(this);
3431         set_control(slow_ctl);
3432         uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
3433       }
3434       replace_in_map(not_null_obj, exact_obj);
3435       obj = exact_obj;
3436     }
3437   } else {
3438     if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3439       Node* exact_obj = null_assert(obj);
3440       replace_in_map(obj, exact_obj);
3441       obj = exact_obj;
3442     }
3443   }
3444   return obj;
3445 }
3446 
3447 //-------------------------------gen_instanceof--------------------------------
3448 // Generate an instance-of idiom.  Used by both the instance-of bytecode
3449 // and the reflective instance-of call.
3450 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
3451   kill_dead_locals();           // Benefit all the uncommon traps
3452   assert( !stopped(), "dead parse path should be checked in callers" );
3453   assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
3454          "must check for not-null not-dead klass in callers");
3455 
3456   // Make the merge point
3457   enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
3458   RegionNode* region = new RegionNode(PATH_LIMIT);
3459   Node*       phi    = new PhiNode(region, TypeInt::BOOL);
3460   C->set_has_split_ifs(true); // Has chance for split-if optimization
3461 
3462   ciProfileData* data = nullptr;
3463   if (java_bc() == Bytecodes::_instanceof) {  // Only for the bytecode
3464     data = method()->method_data()->bci_to_data(bci());
3465   }
3466   bool speculative_not_null = false;
3467   bool never_see_null = (ProfileDynamicTypes  // aggressive use of profile
3468                          && seems_never_null(obj, data, speculative_not_null));
3469 
3470   // Null check; get casted pointer; set region slot 3
3471   Node* null_ctl = top();
3472   Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3473 
3474   // If not_null_obj is dead, only null-path is taken
3475   if (stopped()) {              // Doing instance-of on a null?
3476     set_control(null_ctl);
3477     return intcon(0);
3478   }
3479   region->init_req(_null_path, null_ctl);
3480   phi   ->init_req(_null_path, intcon(0)); // Set null path value
3481   if (null_ctl == top()) {
3482     // Do this eagerly, so that pattern matches like is_diamond_phi
3483     // will work even during parsing.
3484     assert(_null_path == PATH_LIMIT-1, "delete last");
3485     region->del_req(_null_path);
3486     phi   ->del_req(_null_path);
3487   }
3488 
3489   // Do we know the type check always succeed?
3490   bool known_statically = false;
3491   if (_gvn.type(superklass)->singleton()) {
3492     const TypeKlassPtr* superk = _gvn.type(superklass)->is_klassptr();
3493     const TypeKlassPtr* subk = _gvn.type(obj)->is_oopptr()->as_klass_type();
3494     if (subk != nullptr && subk->is_loaded()) {
3495       int static_res = C->static_subtype_check(superk, subk);
3496       known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3497     }
3498   }
3499 
3500   if (!known_statically) {
3501     const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3502     // We may not have profiling here or it may not help us. If we
3503     // have a speculative type use it to perform an exact cast.
3504     ciKlass* spec_obj_type = obj_type->speculative_type();
3505     if (spec_obj_type != nullptr || (ProfileDynamicTypes && data != nullptr)) {
3506       Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, nullptr, spec_obj_type, safe_for_replace);
3507       if (stopped()) {            // Profile disagrees with this path.
3508         set_control(null_ctl);    // Null is the only remaining possibility.
3509         return intcon(0);
3510       }
3511       if (cast_obj != nullptr) {
3512         not_null_obj = cast_obj;
3513       }
3514     }
3515   }
3516 
3517   // Generate the subtype check
3518   Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass);
3519 
3520   // Plug in the success path to the general merge in slot 1.
3521   region->init_req(_obj_path, control());
3522   phi   ->init_req(_obj_path, intcon(1));
3523 
3524   // Plug in the failing path to the general merge in slot 2.
3525   region->init_req(_fail_path, not_subtype_ctrl);
3526   phi   ->init_req(_fail_path, intcon(0));
3527 
3528   // Return final merged results
3529   set_control( _gvn.transform(region) );
3530   record_for_igvn(region);
3531 
3532   // If we know the type check always succeeds then we don't use the
3533   // profiling data at this bytecode. Don't lose it, feed it to the
3534   // type system as a speculative type.
3535   if (safe_for_replace) {
3536     Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3537     replace_in_map(obj, casted_obj);
3538   }
3539 
3540   return _gvn.transform(phi);
3541 }
3542 
3543 //-------------------------------gen_checkcast---------------------------------
3544 // Generate a checkcast idiom.  Used by both the checkcast bytecode and the
3545 // array store bytecode.  Stack must be as-if BEFORE doing the bytecode so the
3546 // uncommon-trap paths work.  Adjust stack after this call.
3547 // If failure_control is supplied and not null, it is filled in with
3548 // the control edge for the cast failure.  Otherwise, an appropriate
3549 // uncommon trap or exception is thrown.
3550 Node* GraphKit::gen_checkcast(Node* obj, Node* superklass, Node* *failure_control, bool null_free, bool maybe_larval) {
3551   kill_dead_locals();           // Benefit all the uncommon traps
3552   const TypeKlassPtr* klass_ptr_type = _gvn.type(superklass)->is_klassptr();
3553   const Type* obj_type = _gvn.type(obj);
3554   if (obj_type->is_inlinetypeptr() && !obj_type->maybe_null() && klass_ptr_type->klass_is_exact() && obj_type->inline_klass() == klass_ptr_type->exact_klass(true)) {
3555     // Special case: larval inline objects must not be scalarized. They are also generally not
3556     // allowed to participate in most operations except as the first operand of putfield, or as an
3557     // argument to a constructor invocation with it being a receiver, Unsafe::putXXX with it being
3558     // the first argument, or Unsafe::finishPrivateBuffer. This allows us to aggressively scalarize
3559     // value objects in all other places. This special case comes from the limitation of the Java
3560     // language, Unsafe::makePrivateBuffer returns an Object that is checkcast-ed to the concrete
3561     // value type. We must do this first because C->static_subtype_check may do nothing when
3562     // StressReflectiveCode is set.
3563     return obj;
3564   }
3565 
3566   // Else it must be a non-larval object
3567   obj = cast_to_non_larval(obj);
3568 
3569   const TypeKlassPtr* improved_klass_ptr_type = klass_ptr_type->try_improve();
3570   const TypeOopPtr* toop = improved_klass_ptr_type->cast_to_exactness(false)->as_instance_type();
3571   bool safe_for_replace = (failure_control == nullptr);
3572   assert(!null_free || toop->can_be_inline_type(), "must be an inline type pointer");
3573 
3574   // Fast cutout:  Check the case that the cast is vacuously true.
3575   // This detects the common cases where the test will short-circuit
3576   // away completely.  We do this before we perform the null check,
3577   // because if the test is going to turn into zero code, we don't
3578   // want a residual null check left around.  (Causes a slowdown,
3579   // for example, in some objArray manipulations, such as a[i]=a[j].)
3580   if (improved_klass_ptr_type->singleton()) {
3581     const TypeKlassPtr* kptr = nullptr;
3582     if (obj_type->isa_oop_ptr()) {
3583       kptr = obj_type->is_oopptr()->as_klass_type();
3584     } else if (obj->is_InlineType()) {
3585       ciInlineKlass* vk = obj_type->inline_klass();
3586       kptr = TypeInstKlassPtr::make(TypePtr::NotNull, vk, Type::Offset(0));
3587     }
3588 
3589     if (kptr != nullptr) {
3590       switch (C->static_subtype_check(improved_klass_ptr_type, kptr)) {
3591       case Compile::SSC_always_true:
3592         // If we know the type check always succeed then we don't use
3593         // the profiling data at this bytecode. Don't lose it, feed it
3594         // to the type system as a speculative type.
3595         obj = record_profiled_receiver_for_speculation(obj);
3596         if (null_free) {
3597           assert(safe_for_replace, "must be");
3598           obj = null_check(obj);
3599         }
3600         assert(stopped() || !toop->is_inlinetypeptr() || obj->is_InlineType(), "should have been scalarized");
3601         return obj;
3602       case Compile::SSC_always_false:
3603         if (null_free) {
3604           assert(safe_for_replace, "must be");
3605           obj = null_check(obj);
3606         }
3607         // It needs a null check because a null will *pass* the cast check.
3608         if (obj_type->isa_oopptr() != nullptr && !obj_type->is_oopptr()->maybe_null()) {
3609           bool is_aastore = (java_bc() == Bytecodes::_aastore);
3610           Deoptimization::DeoptReason reason = is_aastore ?
3611             Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3612           builtin_throw(reason);
3613           return top();
3614         } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3615           return null_assert(obj);
3616         }
3617         break; // Fall through to full check
3618       default:
3619         break;
3620       }
3621     }
3622   }
3623 
3624   ciProfileData* data = nullptr;
3625   if (failure_control == nullptr) {        // use MDO in regular case only
3626     assert(java_bc() == Bytecodes::_aastore ||
3627            java_bc() == Bytecodes::_checkcast,
3628            "interpreter profiles type checks only for these BCs");
3629     if (method()->method_data()->is_mature()) {
3630       data = method()->method_data()->bci_to_data(bci());
3631     }
3632   }
3633 
3634   // Make the merge point
3635   enum { _obj_path = 1, _null_path, PATH_LIMIT };
3636   RegionNode* region = new RegionNode(PATH_LIMIT);
3637   Node*       phi    = new PhiNode(region, toop);
3638   _gvn.set_type(region, Type::CONTROL);
3639   _gvn.set_type(phi, toop);
3640 
3641   C->set_has_split_ifs(true); // Has chance for split-if optimization
3642 
3643   // Use null-cast information if it is available
3644   bool speculative_not_null = false;
3645   bool never_see_null = ((failure_control == nullptr)  // regular case only
3646                          && seems_never_null(obj, data, speculative_not_null));
3647 
3648   if (obj->is_InlineType()) {
3649     // Re-execute if buffering during triggers deoptimization
3650     PreserveReexecuteState preexecs(this);
3651     jvms()->set_should_reexecute(true);
3652     obj = obj->as_InlineType()->buffer(this, safe_for_replace);
3653   }
3654 
3655   // Null check; get casted pointer; set region slot 3
3656   Node* null_ctl = top();
3657   Node* not_null_obj = nullptr;
3658   if (null_free) {
3659     assert(safe_for_replace, "must be");
3660     not_null_obj = null_check(obj);
3661   } else {
3662     not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3663   }
3664 
3665   // If not_null_obj is dead, only null-path is taken
3666   if (stopped()) {              // Doing instance-of on a null?
3667     set_control(null_ctl);
3668     if (toop->is_inlinetypeptr()) {
3669       return InlineTypeNode::make_null(_gvn, toop->inline_klass());
3670     }
3671     return null();
3672   }
3673   region->init_req(_null_path, null_ctl);
3674   phi   ->init_req(_null_path, null());  // Set null path value
3675   if (null_ctl == top()) {
3676     // Do this eagerly, so that pattern matches like is_diamond_phi
3677     // will work even during parsing.
3678     assert(_null_path == PATH_LIMIT-1, "delete last");
3679     region->del_req(_null_path);
3680     phi   ->del_req(_null_path);
3681   }
3682 
3683   Node* cast_obj = nullptr;
3684   if (improved_klass_ptr_type->klass_is_exact()) {
3685     // The following optimization tries to statically cast the speculative type of the object
3686     // (for example obtained during profiling) to the type of the superklass and then do a
3687     // dynamic check that the type of the object is what we expect. To work correctly
3688     // for checkcast and aastore the type of superklass should be exact.
3689     const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3690     // We may not have profiling here or it may not help us. If we have
3691     // a speculative type use it to perform an exact cast.
3692     ciKlass* spec_obj_type = obj_type->speculative_type();
3693     if (spec_obj_type != nullptr || data != nullptr) {
3694       cast_obj = maybe_cast_profiled_receiver(not_null_obj, improved_klass_ptr_type, spec_obj_type, safe_for_replace);
3695       if (cast_obj != nullptr) {
3696         if (failure_control != nullptr) // failure is now impossible
3697           (*failure_control) = top();
3698         // adjust the type of the phi to the exact klass:
3699         phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3700       }
3701     }
3702   }
3703 
3704   if (cast_obj == nullptr) {
3705     // Generate the subtype check
3706     Node* improved_superklass = superklass;
3707     if (improved_klass_ptr_type != klass_ptr_type && improved_klass_ptr_type->singleton()) {
3708       // Only improve the super class for constants which allows subsequent sub type checks to possibly be commoned up.
3709       // The other non-constant cases cannot be improved with a cast node here since they could be folded to top.
3710       // Additionally, the benefit would only be minor in non-constant cases.
3711       improved_superklass = makecon(improved_klass_ptr_type);
3712     }
3713     Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, improved_superklass);
3714     // Plug in success path into the merge
3715     cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3716     // Failure path ends in uncommon trap (or may be dead - failure impossible)
3717     if (failure_control == nullptr) {
3718       if (not_subtype_ctrl != top()) { // If failure is possible
3719         PreserveJVMState pjvms(this);
3720         set_control(not_subtype_ctrl);
3721         bool is_aastore = (java_bc() == Bytecodes::_aastore);
3722         Deoptimization::DeoptReason reason = is_aastore ?
3723           Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3724         builtin_throw(reason);
3725       }
3726     } else {
3727       (*failure_control) = not_subtype_ctrl;
3728     }
3729   }
3730 
3731   region->init_req(_obj_path, control());
3732   phi   ->init_req(_obj_path, cast_obj);
3733 
3734   // A merge of null or Casted-NotNull obj
3735   Node* res = _gvn.transform(phi);
3736 
3737   // Note I do NOT always 'replace_in_map(obj,result)' here.
3738   //  if( tk->klass()->can_be_primary_super()  )
3739     // This means that if I successfully store an Object into an array-of-String
3740     // I 'forget' that the Object is really now known to be a String.  I have to
3741     // do this because we don't have true union types for interfaces - if I store
3742     // a Baz into an array-of-Interface and then tell the optimizer it's an
3743     // Interface, I forget that it's also a Baz and cannot do Baz-like field
3744     // references to it.  FIX THIS WHEN UNION TYPES APPEAR!
3745   //  replace_in_map( obj, res );
3746 
3747   // Return final merged results
3748   set_control( _gvn.transform(region) );
3749   record_for_igvn(region);
3750 
3751   bool not_inline = !toop->can_be_inline_type();
3752   bool not_flat_in_array = !UseArrayFlattening || not_inline || (toop->is_inlinetypeptr() && !toop->inline_klass()->maybe_flat_in_array());
3753   if (Arguments::is_valhalla_enabled() && (not_inline || not_flat_in_array)) {
3754     // Check if obj has been loaded from an array
3755     obj = obj->isa_DecodeN() ? obj->in(1) : obj;
3756     Node* array = nullptr;
3757     if (obj->isa_Load()) {
3758       Node* address = obj->in(MemNode::Address);
3759       if (address->isa_AddP()) {
3760         array = address->as_AddP()->in(AddPNode::Base);
3761       }
3762     } else if (obj->is_Phi()) {
3763       Node* region = obj->in(0);
3764       // TODO make this more robust (see JDK-8231346)
3765       if (region->req() == 3 && region->in(2) != nullptr && region->in(2)->in(0) != nullptr) {
3766         IfNode* iff = region->in(2)->in(0)->isa_If();
3767         if (iff != nullptr) {
3768           iff->is_flat_array_check(&_gvn, &array);
3769         }
3770       }
3771     }
3772     if (array != nullptr) {
3773       const TypeAryPtr* ary_t = _gvn.type(array)->isa_aryptr();
3774       if (ary_t != nullptr) {
3775         if (!ary_t->is_not_null_free() && !ary_t->is_null_free() && not_inline) {
3776           // Casting array element to a non-inline-type, mark array as not null-free.
3777           Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_null_free()));
3778           replace_in_map(array, cast);
3779           array = cast;
3780         }
3781         if (!ary_t->is_not_flat() && !ary_t->is_flat() && not_flat_in_array) {
3782           // Casting array element to a non-flat-in-array type, mark array as not flat.
3783           Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_flat()));
3784           replace_in_map(array, cast);
3785           array = cast;
3786         }
3787       }
3788     }
3789   }
3790 
3791   if (!stopped() && !res->is_InlineType()) {
3792     res = record_profiled_receiver_for_speculation(res);
3793     if (toop->is_inlinetypeptr() && !maybe_larval) {
3794       Node* vt = InlineTypeNode::make_from_oop(this, res, toop->inline_klass());
3795       res = vt;
3796       if (safe_for_replace) {
3797         replace_in_map(obj, vt);
3798         replace_in_map(not_null_obj, vt);
3799         replace_in_map(res, vt);
3800       }
3801     }
3802   }
3803   return res;
3804 }
3805 
3806 Node* GraphKit::mark_word_test(Node* obj, uintptr_t mask_val, bool eq, bool check_lock) {
3807   // Load markword
3808   Node* mark_adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3809   Node* mark = make_load(nullptr, mark_adr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
3810   if (check_lock && !UseCompactObjectHeaders) {
3811     // COH: Locking does not override the markword with a tagged pointer. We can directly read from the markword.
3812     // Check if obj is locked
3813     Node* locked_bit = MakeConX(markWord::unlocked_value);
3814     locked_bit = _gvn.transform(new AndXNode(locked_bit, mark));
3815     Node* cmp = _gvn.transform(new CmpXNode(locked_bit, MakeConX(0)));
3816     Node* is_unlocked = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3817     IfNode* iff = new IfNode(control(), is_unlocked, PROB_MAX, COUNT_UNKNOWN);
3818     _gvn.transform(iff);
3819     Node* locked_region = new RegionNode(3);
3820     Node* mark_phi = new PhiNode(locked_region, TypeX_X);
3821 
3822     // Unlocked: Use bits from mark word
3823     locked_region->init_req(1, _gvn.transform(new IfTrueNode(iff)));
3824     mark_phi->init_req(1, mark);
3825 
3826     // Locked: Load prototype header from klass
3827     set_control(_gvn.transform(new IfFalseNode(iff)));
3828     // Make loads control dependent to make sure they are only executed if array is locked
3829     Node* klass_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
3830     Node* klass = _gvn.transform(LoadKlassNode::make(_gvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
3831     Node* proto_adr = basic_plus_adr(klass, in_bytes(Klass::prototype_header_offset()));
3832     Node* proto = _gvn.transform(LoadNode::make(_gvn, control(), C->immutable_memory(), proto_adr, proto_adr->bottom_type()->is_ptr(), TypeX_X, TypeX_X->basic_type(), MemNode::unordered));
3833 
3834     locked_region->init_req(2, control());
3835     mark_phi->init_req(2, proto);
3836     set_control(_gvn.transform(locked_region));
3837     record_for_igvn(locked_region);
3838 
3839     mark = mark_phi;
3840   }
3841 
3842   // Now check if mark word bits are set
3843   Node* mask = MakeConX(mask_val);
3844   Node* masked = _gvn.transform(new AndXNode(_gvn.transform(mark), mask));
3845   record_for_igvn(masked); // Give it a chance to be optimized out by IGVN
3846   Node* cmp = _gvn.transform(new CmpXNode(masked, mask));
3847   return _gvn.transform(new BoolNode(cmp, eq ? BoolTest::eq : BoolTest::ne));
3848 }
3849 
3850 Node* GraphKit::inline_type_test(Node* obj, bool is_inline) {
3851   return mark_word_test(obj, markWord::inline_type_pattern, is_inline, /* check_lock = */ false);
3852 }
3853 
3854 Node* GraphKit::flat_array_test(Node* array_or_klass, bool flat) {
3855   // We can't use immutable memory here because the mark word is mutable.
3856   // PhaseIdealLoop::move_flat_array_check_out_of_loop will make sure the
3857   // check is moved out of loops (mainly to enable loop unswitching).
3858   Node* cmp = _gvn.transform(new FlatArrayCheckNode(C, memory(Compile::AliasIdxRaw), array_or_klass));
3859   record_for_igvn(cmp); // Give it a chance to be optimized out by IGVN
3860   return _gvn.transform(new BoolNode(cmp, flat ? BoolTest::eq : BoolTest::ne));
3861 }
3862 
3863 Node* GraphKit::null_free_array_test(Node* array, bool null_free) {
3864   return mark_word_test(array, markWord::null_free_array_bit_in_place, null_free);
3865 }
3866 
3867 Node* GraphKit::null_free_atomic_array_test(Node* array, ciInlineKlass* vk) {
3868   assert(vk->has_atomic_layout() || vk->has_non_atomic_layout(), "Can't be null-free and flat");
3869 
3870   // TODO 8350865 Add a stress flag to always access atomic if layout exists?
3871   if (!vk->has_non_atomic_layout()) {
3872     return intcon(1); // Always atomic
3873   } else if (!vk->has_atomic_layout()) {
3874     return intcon(0); // Never atomic
3875   }
3876 
3877   Node* array_klass = load_object_klass(array);
3878   int layout_kind_offset = in_bytes(FlatArrayKlass::layout_kind_offset());
3879   Node* layout_kind_addr = basic_plus_adr(array_klass, array_klass, layout_kind_offset);
3880   Node* layout_kind = make_load(nullptr, layout_kind_addr, TypeInt::INT, T_INT, MemNode::unordered);
3881   Node* cmp = _gvn.transform(new CmpINode(layout_kind, intcon((int)LayoutKind::NULL_FREE_ATOMIC_FLAT)));
3882   return _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3883 }
3884 
3885 // Deoptimize if 'ary' is a null-free inline type array and 'val' is null
3886 Node* GraphKit::inline_array_null_guard(Node* ary, Node* val, int nargs, bool safe_for_replace) {
3887   RegionNode* region = new RegionNode(3);
3888   Node* null_ctl = top();
3889   null_check_oop(val, &null_ctl);
3890   if (null_ctl != top()) {
3891     PreserveJVMState pjvms(this);
3892     set_control(null_ctl);
3893     {
3894       // Deoptimize if null-free array
3895       BuildCutout unless(this, null_free_array_test(ary, /* null_free = */ false), PROB_MAX);
3896       inc_sp(nargs);
3897       uncommon_trap(Deoptimization::Reason_null_check,
3898                     Deoptimization::Action_none);
3899     }
3900     region->init_req(1, control());
3901   }
3902   region->init_req(2, control());
3903   set_control(_gvn.transform(region));
3904   record_for_igvn(region);
3905   if (_gvn.type(val) == TypePtr::NULL_PTR) {
3906     // Since we were just successfully storing null, the array can't be null free.
3907     const TypeAryPtr* ary_t = _gvn.type(ary)->is_aryptr();
3908     ary_t = ary_t->cast_to_not_null_free();
3909     Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, ary_t));
3910     if (safe_for_replace) {
3911       replace_in_map(ary, cast);
3912     }
3913     ary = cast;
3914   }
3915   return ary;
3916 }
3917 
3918 //------------------------------next_monitor-----------------------------------
3919 // What number should be given to the next monitor?
3920 int GraphKit::next_monitor() {
3921   int current = jvms()->monitor_depth()* C->sync_stack_slots();
3922   int next = current + C->sync_stack_slots();
3923   // Keep the toplevel high water mark current:
3924   if (C->fixed_slots() < next)  C->set_fixed_slots(next);
3925   return current;
3926 }
3927 
3928 //------------------------------insert_mem_bar---------------------------------
3929 // Memory barrier to avoid floating things around
3930 // The membar serves as a pinch point between both control and all memory slices.
3931 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3932   MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3933   mb->init_req(TypeFunc::Control, control());
3934   mb->init_req(TypeFunc::Memory,  reset_memory());
3935   Node* membar = _gvn.transform(mb);
3936   set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3937   set_all_memory_call(membar);
3938   return membar;
3939 }
3940 
3941 //-------------------------insert_mem_bar_volatile----------------------------
3942 // Memory barrier to avoid floating things around
3943 // The membar serves as a pinch point between both control and memory(alias_idx).
3944 // If you want to make a pinch point on all memory slices, do not use this
3945 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3946 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3947   // When Parse::do_put_xxx updates a volatile field, it appends a series
3948   // of MemBarVolatile nodes, one for *each* volatile field alias category.
3949   // The first membar is on the same memory slice as the field store opcode.
3950   // This forces the membar to follow the store.  (Bug 6500685 broke this.)
3951   // All the other membars (for other volatile slices, including AliasIdxBot,
3952   // which stands for all unknown volatile slices) are control-dependent
3953   // on the first membar.  This prevents later volatile loads or stores
3954   // from sliding up past the just-emitted store.
3955 
3956   MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3957   mb->set_req(TypeFunc::Control,control());
3958   if (alias_idx == Compile::AliasIdxBot) {
3959     mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3960   } else {
3961     assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3962     mb->set_req(TypeFunc::Memory, memory(alias_idx));
3963   }
3964   Node* membar = _gvn.transform(mb);
3965   set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3966   if (alias_idx == Compile::AliasIdxBot) {
3967     merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3968   } else {
3969     set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3970   }
3971   return membar;
3972 }
3973 
3974 //------------------------------shared_lock------------------------------------
3975 // Emit locking code.
3976 FastLockNode* GraphKit::shared_lock(Node* obj) {
3977   // bci is either a monitorenter bc or InvocationEntryBci
3978   // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3979   assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3980 
3981   if (stopped())                // Dead monitor?
3982     return nullptr;
3983 
3984   assert(dead_locals_are_killed(), "should kill locals before sync. point");
3985 
3986   // Box the stack location
3987   Node* box = new BoxLockNode(next_monitor());
3988   // Check for bailout after new BoxLockNode
3989   if (failing()) { return nullptr; }
3990   box = _gvn.transform(box);
3991   Node* mem = reset_memory();
3992 
3993   FastLockNode * flock = _gvn.transform(new FastLockNode(nullptr, obj, box) )->as_FastLock();
3994 
3995   // Add monitor to debug info for the slow path.  If we block inside the
3996   // slow path and de-opt, we need the monitor hanging around
3997   map()->push_monitor( flock );
3998 
3999   const TypeFunc *tf = LockNode::lock_type();
4000   LockNode *lock = new LockNode(C, tf);
4001 
4002   lock->init_req( TypeFunc::Control, control() );
4003   lock->init_req( TypeFunc::Memory , mem );
4004   lock->init_req( TypeFunc::I_O    , top() )     ;   // does no i/o
4005   lock->init_req( TypeFunc::FramePtr, frameptr() );
4006   lock->init_req( TypeFunc::ReturnAdr, top() );
4007 
4008   lock->init_req(TypeFunc::Parms + 0, obj);
4009   lock->init_req(TypeFunc::Parms + 1, box);
4010   lock->init_req(TypeFunc::Parms + 2, flock);
4011   add_safepoint_edges(lock);
4012 
4013   lock = _gvn.transform( lock )->as_Lock();
4014 
4015   // lock has no side-effects, sets few values
4016   set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
4017 
4018   insert_mem_bar(Op_MemBarAcquireLock);
4019 
4020   // Add this to the worklist so that the lock can be eliminated
4021   record_for_igvn(lock);
4022 
4023 #ifndef PRODUCT
4024   if (PrintLockStatistics) {
4025     // Update the counter for this lock.  Don't bother using an atomic
4026     // operation since we don't require absolute accuracy.
4027     lock->create_lock_counter(map()->jvms());
4028     increment_counter(lock->counter()->addr());
4029   }
4030 #endif
4031 
4032   return flock;
4033 }
4034 
4035 
4036 //------------------------------shared_unlock----------------------------------
4037 // Emit unlocking code.
4038 void GraphKit::shared_unlock(Node* box, Node* obj) {
4039   // bci is either a monitorenter bc or InvocationEntryBci
4040   // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
4041   assert(SynchronizationEntryBCI == InvocationEntryBci, "");
4042 
4043   if (stopped()) {               // Dead monitor?
4044     map()->pop_monitor();        // Kill monitor from debug info
4045     return;
4046   }
4047   assert(!obj->is_InlineType(), "should not unlock on inline type");
4048 
4049   // Memory barrier to avoid floating things down past the locked region
4050   insert_mem_bar(Op_MemBarReleaseLock);
4051 
4052   const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
4053   UnlockNode *unlock = new UnlockNode(C, tf);
4054 #ifdef ASSERT
4055   unlock->set_dbg_jvms(sync_jvms());
4056 #endif
4057   uint raw_idx = Compile::AliasIdxRaw;
4058   unlock->init_req( TypeFunc::Control, control() );
4059   unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
4060   unlock->init_req( TypeFunc::I_O    , top() )     ;   // does no i/o
4061   unlock->init_req( TypeFunc::FramePtr, frameptr() );
4062   unlock->init_req( TypeFunc::ReturnAdr, top() );
4063 
4064   unlock->init_req(TypeFunc::Parms + 0, obj);
4065   unlock->init_req(TypeFunc::Parms + 1, box);
4066   unlock = _gvn.transform(unlock)->as_Unlock();
4067 
4068   Node* mem = reset_memory();
4069 
4070   // unlock has no side-effects, sets few values
4071   set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
4072 
4073   // Kill monitor from debug info
4074   map()->pop_monitor( );
4075 }
4076 
4077 //-------------------------------get_layout_helper-----------------------------
4078 // If the given klass is a constant or known to be an array,
4079 // fetch the constant layout helper value into constant_value
4080 // and return null.  Otherwise, load the non-constant
4081 // layout helper value, and return the node which represents it.
4082 // This two-faced routine is useful because allocation sites
4083 // almost always feature constant types.
4084 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
4085   const TypeKlassPtr* klass_t = _gvn.type(klass_node)->isa_klassptr();
4086   if (!StressReflectiveCode && klass_t != nullptr) {
4087     bool xklass = klass_t->klass_is_exact();
4088     bool can_be_flat = false;
4089     const TypeAryPtr* ary_type = klass_t->as_instance_type()->isa_aryptr();
4090     if (UseArrayFlattening && !xklass && ary_type != nullptr && !ary_type->is_null_free()) {
4091       // Don't constant fold if the runtime type might be a flat array but the static type is not.
4092       const TypeOopPtr* elem = ary_type->elem()->make_oopptr();
4093       can_be_flat = ary_type->can_be_inline_array() && (!elem->is_inlinetypeptr() || elem->inline_klass()->maybe_flat_in_array());
4094     }
4095     if (!can_be_flat && (xklass || (klass_t->isa_aryklassptr() && klass_t->is_aryklassptr()->elem() != Type::BOTTOM))) {
4096       jint lhelper;
4097       if (klass_t->is_flat()) {
4098         lhelper = ary_type->flat_layout_helper();
4099       } else if (klass_t->isa_aryklassptr()) {
4100         BasicType elem = ary_type->elem()->array_element_basic_type();
4101         if (is_reference_type(elem, true)) {
4102           elem = T_OBJECT;
4103         }
4104         lhelper = Klass::array_layout_helper(elem);
4105       } else {
4106         lhelper = klass_t->is_instklassptr()->exact_klass()->layout_helper();
4107       }
4108       if (lhelper != Klass::_lh_neutral_value) {
4109         constant_value = lhelper;
4110         return (Node*) nullptr;
4111       }
4112     }
4113   }
4114   constant_value = Klass::_lh_neutral_value;  // put in a known value
4115   Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
4116   return make_load(nullptr, lhp, TypeInt::INT, T_INT, MemNode::unordered);
4117 }
4118 
4119 // We just put in an allocate/initialize with a big raw-memory effect.
4120 // Hook selected additional alias categories on the initialization.
4121 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
4122                                 MergeMemNode* init_in_merge,
4123                                 Node* init_out_raw) {
4124   DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
4125   assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
4126 
4127   Node* prevmem = kit.memory(alias_idx);
4128   init_in_merge->set_memory_at(alias_idx, prevmem);
4129   if (init_out_raw != nullptr) {
4130     kit.set_memory(init_out_raw, alias_idx);
4131   }
4132 }
4133 
4134 //---------------------------set_output_for_allocation-------------------------
4135 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
4136                                           const TypeOopPtr* oop_type,
4137                                           bool deoptimize_on_exception) {
4138   int rawidx = Compile::AliasIdxRaw;
4139   alloc->set_req( TypeFunc::FramePtr, frameptr() );
4140   add_safepoint_edges(alloc);
4141   Node* allocx = _gvn.transform(alloc);
4142   set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
4143   // create memory projection for i_o
4144   set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
4145   make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
4146 
4147   // create a memory projection as for the normal control path
4148   Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
4149   set_memory(malloc, rawidx);
4150 
4151   // a normal slow-call doesn't change i_o, but an allocation does
4152   // we create a separate i_o projection for the normal control path
4153   set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
4154   Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
4155 
4156   // put in an initialization barrier
4157   InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
4158                                                  rawoop)->as_Initialize();
4159   assert(alloc->initialization() == init,  "2-way macro link must work");
4160   assert(init ->allocation()     == alloc, "2-way macro link must work");
4161   {
4162     // Extract memory strands which may participate in the new object's
4163     // initialization, and source them from the new InitializeNode.
4164     // This will allow us to observe initializations when they occur,
4165     // and link them properly (as a group) to the InitializeNode.
4166     assert(init->in(InitializeNode::Memory) == malloc, "");
4167     MergeMemNode* minit_in = MergeMemNode::make(malloc);
4168     init->set_req(InitializeNode::Memory, minit_in);
4169     record_for_igvn(minit_in); // fold it up later, if possible
4170     _gvn.set_type(minit_in, Type::MEMORY);
4171     Node* minit_out = memory(rawidx);
4172     assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
4173     int mark_idx = C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes()));
4174     // Add an edge in the MergeMem for the header fields so an access to one of those has correct memory state.
4175     // Use one NarrowMemProjNode per slice to properly record the adr type of each slice. The Initialize node will have
4176     // multiple projections as a result.
4177     set_memory(_gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(mark_idx))), mark_idx);
4178     int klass_idx = C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes()));
4179     set_memory(_gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(klass_idx))), klass_idx);
4180     if (oop_type->isa_aryptr()) {
4181       // Initially all flat array accesses share a single slice
4182       // but that changes after parsing. Prepare the memory graph so
4183       // it can optimize flat array accesses properly once they
4184       // don't share a single slice.
4185       assert(C->flat_accesses_share_alias(), "should be set at parse time");
4186       const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
4187       int            elemidx  = C->get_alias_index(telemref);
4188       hook_memory_on_init(*this, elemidx, minit_in, _gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(elemidx))));
4189     } else if (oop_type->isa_instptr()) {
4190       set_memory(minit_out, C->get_alias_index(oop_type)); // mark word
4191       ciInstanceKlass* ik = oop_type->is_instptr()->instance_klass();
4192       for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
4193         ciField* field = ik->nonstatic_field_at(i);
4194         if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize)
4195           continue;  // do not bother to track really large numbers of fields
4196         // Find (or create) the alias category for this field:
4197         int fieldidx = C->alias_type(field)->index();
4198         hook_memory_on_init(*this, fieldidx, minit_in, _gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(fieldidx))));
4199       }
4200     }
4201   }
4202 
4203   // Cast raw oop to the real thing...
4204   Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
4205   javaoop = _gvn.transform(javaoop);
4206   C->set_recent_alloc(control(), javaoop);
4207   assert(just_allocated_object(control()) == javaoop, "just allocated");
4208 
4209 #ifdef ASSERT
4210   { // Verify that the AllocateNode::Ideal_allocation recognizers work:
4211     assert(AllocateNode::Ideal_allocation(rawoop) == alloc,
4212            "Ideal_allocation works");
4213     assert(AllocateNode::Ideal_allocation(javaoop) == alloc,
4214            "Ideal_allocation works");
4215     if (alloc->is_AllocateArray()) {
4216       assert(AllocateArrayNode::Ideal_array_allocation(rawoop) == alloc->as_AllocateArray(),
4217              "Ideal_allocation works");
4218       assert(AllocateArrayNode::Ideal_array_allocation(javaoop) == alloc->as_AllocateArray(),
4219              "Ideal_allocation works");
4220     } else {
4221       assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
4222     }
4223   }
4224 #endif //ASSERT
4225 
4226   return javaoop;
4227 }
4228 
4229 //---------------------------new_instance--------------------------------------
4230 // This routine takes a klass_node which may be constant (for a static type)
4231 // or may be non-constant (for reflective code).  It will work equally well
4232 // for either, and the graph will fold nicely if the optimizer later reduces
4233 // the type to a constant.
4234 // The optional arguments are for specialized use by intrinsics:
4235 //  - If 'extra_slow_test' if not null is an extra condition for the slow-path.
4236 //  - If 'return_size_val', report the total object size to the caller.
4237 //  - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
4238 Node* GraphKit::new_instance(Node* klass_node,
4239                              Node* extra_slow_test,
4240                              Node* *return_size_val,
4241                              bool deoptimize_on_exception,
4242                              InlineTypeNode* inline_type_node) {
4243   // Compute size in doublewords
4244   // The size is always an integral number of doublewords, represented
4245   // as a positive bytewise size stored in the klass's layout_helper.
4246   // The layout_helper also encodes (in a low bit) the need for a slow path.
4247   jint  layout_con = Klass::_lh_neutral_value;
4248   Node* layout_val = get_layout_helper(klass_node, layout_con);
4249   bool  layout_is_con = (layout_val == nullptr);
4250 
4251   if (extra_slow_test == nullptr)  extra_slow_test = intcon(0);
4252   // Generate the initial go-slow test.  It's either ALWAYS (return a
4253   // Node for 1) or NEVER (return a null) or perhaps (in the reflective
4254   // case) a computed value derived from the layout_helper.
4255   Node* initial_slow_test = nullptr;
4256   if (layout_is_con) {
4257     assert(!StressReflectiveCode, "stress mode does not use these paths");
4258     bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
4259     initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
4260   } else {   // reflective case
4261     // This reflective path is used by Unsafe.allocateInstance.
4262     // (It may be stress-tested by specifying StressReflectiveCode.)
4263     // Basically, we want to get into the VM is there's an illegal argument.
4264     Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
4265     initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
4266     if (extra_slow_test != intcon(0)) {
4267       initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
4268     }
4269     // (Macro-expander will further convert this to a Bool, if necessary.)
4270   }
4271 
4272   // Find the size in bytes.  This is easy; it's the layout_helper.
4273   // The size value must be valid even if the slow path is taken.
4274   Node* size = nullptr;
4275   if (layout_is_con) {
4276     size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
4277   } else {   // reflective case
4278     // This reflective path is used by clone and Unsafe.allocateInstance.
4279     size = ConvI2X(layout_val);
4280 
4281     // Clear the low bits to extract layout_helper_size_in_bytes:
4282     assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
4283     Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
4284     size = _gvn.transform( new AndXNode(size, mask) );
4285   }
4286   if (return_size_val != nullptr) {
4287     (*return_size_val) = size;
4288   }
4289 
4290   // This is a precise notnull oop of the klass.
4291   // (Actually, it need not be precise if this is a reflective allocation.)
4292   // It's what we cast the result to.
4293   const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
4294   if (!tklass)  tklass = TypeInstKlassPtr::OBJECT;
4295   const TypeOopPtr* oop_type = tklass->as_instance_type();
4296 
4297   // Now generate allocation code
4298 
4299   // The entire memory state is needed for slow path of the allocation
4300   // since GC and deoptimization can happen.
4301   Node *mem = reset_memory();
4302   set_all_memory(mem); // Create new memory state
4303 
4304   AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
4305                                          control(), mem, i_o(),
4306                                          size, klass_node,
4307                                          initial_slow_test, inline_type_node);
4308 
4309   return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
4310 }
4311 
4312 //-------------------------------new_array-------------------------------------
4313 // helper for newarray and anewarray
4314 // The 'length' parameter is (obviously) the length of the array.
4315 // The optional arguments are for specialized use by intrinsics:
4316 //  - If 'return_size_val', report the non-padded array size (sum of header size
4317 //    and array body) to the caller.
4318 //  - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
4319 Node* GraphKit::new_array(Node* klass_node,     // array klass (maybe variable)
4320                           Node* length,         // number of array elements
4321                           int   nargs,          // number of arguments to push back for uncommon trap
4322                           Node* *return_size_val,
4323                           bool deoptimize_on_exception,
4324                           Node* init_val) {
4325   jint  layout_con = Klass::_lh_neutral_value;
4326   Node* layout_val = get_layout_helper(klass_node, layout_con);
4327   bool  layout_is_con = (layout_val == nullptr);
4328 
4329   if (!layout_is_con && !StressReflectiveCode &&
4330       !too_many_traps(Deoptimization::Reason_class_check)) {
4331     // This is a reflective array creation site.
4332     // Optimistically assume that it is a subtype of Object[],
4333     // so that we can fold up all the address arithmetic.
4334     layout_con = Klass::array_layout_helper(T_OBJECT);
4335     Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
4336     Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
4337     { BuildCutout unless(this, bol_lh, PROB_MAX);
4338       inc_sp(nargs);
4339       uncommon_trap(Deoptimization::Reason_class_check,
4340                     Deoptimization::Action_maybe_recompile);
4341     }
4342     layout_val = nullptr;
4343     layout_is_con = true;
4344   }
4345 
4346   // Generate the initial go-slow test.  Make sure we do not overflow
4347   // if length is huge (near 2Gig) or negative!  We do not need
4348   // exact double-words here, just a close approximation of needed
4349   // double-words.  We can't add any offset or rounding bits, lest we
4350   // take a size -1 of bytes and make it positive.  Use an unsigned
4351   // compare, so negative sizes look hugely positive.
4352   int fast_size_limit = FastAllocateSizeLimit;
4353   if (layout_is_con) {
4354     assert(!StressReflectiveCode, "stress mode does not use these paths");
4355     // Increase the size limit if we have exact knowledge of array type.
4356     int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
4357     fast_size_limit <<= MAX2(LogBytesPerLong - log2_esize, 0);
4358   }
4359 
4360   Node* initial_slow_cmp  = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
4361   Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
4362 
4363   // --- Size Computation ---
4364   // array_size = round_to_heap(array_header + (length << elem_shift));
4365   // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
4366   // and align_to(x, y) == ((x + y-1) & ~(y-1))
4367   // The rounding mask is strength-reduced, if possible.
4368   int round_mask = MinObjAlignmentInBytes - 1;
4369   Node* header_size = nullptr;
4370   // (T_BYTE has the weakest alignment and size restrictions...)
4371   if (layout_is_con) {
4372     int       hsize  = Klass::layout_helper_header_size(layout_con);
4373     int       eshift = Klass::layout_helper_log2_element_size(layout_con);
4374     bool is_flat_array = Klass::layout_helper_is_flatArray(layout_con);
4375     if ((round_mask & ~right_n_bits(eshift)) == 0)
4376       round_mask = 0;  // strength-reduce it if it goes away completely
4377     assert(is_flat_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
4378     int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
4379     assert(header_size_min <= hsize, "generic minimum is smallest");
4380     header_size = intcon(hsize);
4381   } else {
4382     Node* hss   = intcon(Klass::_lh_header_size_shift);
4383     Node* hsm   = intcon(Klass::_lh_header_size_mask);
4384     header_size = _gvn.transform(new URShiftINode(layout_val, hss));
4385     header_size = _gvn.transform(new AndINode(header_size, hsm));
4386   }
4387 
4388   Node* elem_shift = nullptr;
4389   if (layout_is_con) {
4390     int eshift = Klass::layout_helper_log2_element_size(layout_con);
4391     if (eshift != 0)
4392       elem_shift = intcon(eshift);
4393   } else {
4394     // There is no need to mask or shift this value.
4395     // The semantics of LShiftINode include an implicit mask to 0x1F.
4396     assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
4397     elem_shift = layout_val;
4398   }
4399 
4400   // Transition to native address size for all offset calculations:
4401   Node* lengthx = ConvI2X(length);
4402   Node* headerx = ConvI2X(header_size);
4403 #ifdef _LP64
4404   { const TypeInt* tilen = _gvn.find_int_type(length);
4405     if (tilen != nullptr && tilen->_lo < 0) {
4406       // Add a manual constraint to a positive range.  Cf. array_element_address.
4407       jint size_max = fast_size_limit;
4408       if (size_max > tilen->_hi && tilen->_hi >= 0) {
4409         size_max = tilen->_hi;
4410       }
4411       const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
4412 
4413       // Only do a narrow I2L conversion if the range check passed.
4414       IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
4415       _gvn.transform(iff);
4416       RegionNode* region = new RegionNode(3);
4417       _gvn.set_type(region, Type::CONTROL);
4418       lengthx = new PhiNode(region, TypeLong::LONG);
4419       _gvn.set_type(lengthx, TypeLong::LONG);
4420 
4421       // Range check passed. Use ConvI2L node with narrow type.
4422       Node* passed = IfFalse(iff);
4423       region->init_req(1, passed);
4424       // Make I2L conversion control dependent to prevent it from
4425       // floating above the range check during loop optimizations.
4426       lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
4427 
4428       // Range check failed. Use ConvI2L with wide type because length may be invalid.
4429       region->init_req(2, IfTrue(iff));
4430       lengthx->init_req(2, ConvI2X(length));
4431 
4432       set_control(region);
4433       record_for_igvn(region);
4434       record_for_igvn(lengthx);
4435     }
4436   }
4437 #endif
4438 
4439   // Combine header size and body size for the array copy part, then align (if
4440   // necessary) for the allocation part. This computation cannot overflow,
4441   // because it is used only in two places, one where the length is sharply
4442   // limited, and the other after a successful allocation.
4443   Node* abody = lengthx;
4444   if (elem_shift != nullptr) {
4445     abody = _gvn.transform(new LShiftXNode(lengthx, elem_shift));
4446   }
4447   Node* non_rounded_size = _gvn.transform(new AddXNode(headerx, abody));
4448 
4449   if (return_size_val != nullptr) {
4450     // This is the size
4451     (*return_size_val) = non_rounded_size;
4452   }
4453 
4454   Node* size = non_rounded_size;
4455   if (round_mask != 0) {
4456     Node* mask1 = MakeConX(round_mask);
4457     size = _gvn.transform(new AddXNode(size, mask1));
4458     Node* mask2 = MakeConX(~round_mask);
4459     size = _gvn.transform(new AndXNode(size, mask2));
4460   }
4461   // else if round_mask == 0, the size computation is self-rounding
4462 
4463   // Now generate allocation code
4464 
4465   // The entire memory state is needed for slow path of the allocation
4466   // since GC and deoptimization can happen.
4467   Node *mem = reset_memory();
4468   set_all_memory(mem); // Create new memory state
4469 
4470   if (initial_slow_test->is_Bool()) {
4471     // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
4472     initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
4473   }
4474 
4475   const TypeKlassPtr* ary_klass = _gvn.type(klass_node)->isa_klassptr();
4476   const TypeOopPtr* ary_type = ary_klass->as_instance_type();
4477 
4478   Node* raw_init_value = nullptr;
4479   if (init_val != nullptr) {
4480     // TODO 8350865 Fast non-zero init not implemented yet for flat, null-free arrays
4481     if (ary_type->is_flat()) {
4482       initial_slow_test = intcon(1);
4483     }
4484 
4485     if (UseCompressedOops) {
4486       // With compressed oops, the 64-bit init value is built from two 32-bit compressed oops
4487       init_val = _gvn.transform(new EncodePNode(init_val, init_val->bottom_type()->make_narrowoop()));
4488       Node* lower = _gvn.transform(new CastP2XNode(control(), init_val));
4489       Node* upper = _gvn.transform(new LShiftLNode(lower, intcon(32)));
4490       raw_init_value = _gvn.transform(new OrLNode(lower, upper));
4491     } else {
4492       raw_init_value = _gvn.transform(new CastP2XNode(control(), init_val));
4493     }
4494   }
4495 
4496   Node* valid_length_test = _gvn.intcon(1);
4497   if (ary_type->isa_aryptr()) {
4498     BasicType bt = ary_type->isa_aryptr()->elem()->array_element_basic_type();
4499     jint max = TypeAryPtr::max_array_length(bt);
4500     Node* valid_length_cmp  = _gvn.transform(new CmpUNode(length, intcon(max)));
4501     valid_length_test = _gvn.transform(new BoolNode(valid_length_cmp, BoolTest::le));
4502   }
4503 
4504   // Create the AllocateArrayNode and its result projections
4505   AllocateArrayNode* alloc
4506     = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
4507                             control(), mem, i_o(),
4508                             size, klass_node,
4509                             initial_slow_test,
4510                             length, valid_length_test,
4511                             init_val, raw_init_value);
4512   // Cast to correct type.  Note that the klass_node may be constant or not,
4513   // and in the latter case the actual array type will be inexact also.
4514   // (This happens via a non-constant argument to inline_native_newArray.)
4515   // In any case, the value of klass_node provides the desired array type.
4516   const TypeInt* length_type = _gvn.find_int_type(length);
4517   if (ary_type->isa_aryptr() && length_type != nullptr) {
4518     // Try to get a better type than POS for the size
4519     ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4520   }
4521 
4522   Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
4523 
4524   array_ideal_length(alloc, ary_type, true);
4525   return javaoop;
4526 }
4527 
4528 // The following "Ideal_foo" functions are placed here because they recognize
4529 // the graph shapes created by the functions immediately above.
4530 
4531 //---------------------------Ideal_allocation----------------------------------
4532 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
4533 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr) {
4534   if (ptr == nullptr) {     // reduce dumb test in callers
4535     return nullptr;
4536   }
4537 
4538   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4539   ptr = bs->step_over_gc_barrier(ptr);
4540 
4541   if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
4542     ptr = ptr->in(1);
4543     if (ptr == nullptr) return nullptr;
4544   }
4545   // Return null for allocations with several casts:
4546   //   j.l.reflect.Array.newInstance(jobject, jint)
4547   //   Object.clone()
4548   // to keep more precise type from last cast.
4549   if (ptr->is_Proj()) {
4550     Node* allo = ptr->in(0);
4551     if (allo != nullptr && allo->is_Allocate()) {
4552       return allo->as_Allocate();
4553     }
4554   }
4555   // Report failure to match.
4556   return nullptr;
4557 }
4558 
4559 // Fancy version which also strips off an offset (and reports it to caller).
4560 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseValues* phase,
4561                                              intptr_t& offset) {
4562   Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
4563   if (base == nullptr)  return nullptr;
4564   return Ideal_allocation(base);
4565 }
4566 
4567 // Trace Initialize <- Proj[Parm] <- Allocate
4568 AllocateNode* InitializeNode::allocation() {
4569   Node* rawoop = in(InitializeNode::RawAddress);
4570   if (rawoop->is_Proj()) {
4571     Node* alloc = rawoop->in(0);
4572     if (alloc->is_Allocate()) {
4573       return alloc->as_Allocate();
4574     }
4575   }
4576   return nullptr;
4577 }
4578 
4579 // Trace Allocate -> Proj[Parm] -> Initialize
4580 InitializeNode* AllocateNode::initialization() {
4581   ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress);
4582   if (rawoop == nullptr)  return nullptr;
4583   for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
4584     Node* init = rawoop->fast_out(i);
4585     if (init->is_Initialize()) {
4586       assert(init->as_Initialize()->allocation() == this, "2-way link");
4587       return init->as_Initialize();
4588     }
4589   }
4590   return nullptr;
4591 }
4592 
4593 // Add a Parse Predicate with an uncommon trap on the failing/false path. Normal control will continue on the true path.
4594 void GraphKit::add_parse_predicate(Deoptimization::DeoptReason reason, const int nargs) {
4595   // Too many traps seen?
4596   if (too_many_traps(reason)) {
4597 #ifdef ASSERT
4598     if (TraceLoopPredicate) {
4599       int tc = C->trap_count(reason);
4600       tty->print("too many traps=%s tcount=%d in ",
4601                     Deoptimization::trap_reason_name(reason), tc);
4602       method()->print(); // which method has too many predicate traps
4603       tty->cr();
4604     }
4605 #endif
4606     // We cannot afford to take more traps here,
4607     // do not generate Parse Predicate.
4608     return;
4609   }
4610 
4611   ParsePredicateNode* parse_predicate = new ParsePredicateNode(control(), reason, &_gvn);
4612   _gvn.set_type(parse_predicate, parse_predicate->Value(&_gvn));
4613   Node* if_false = _gvn.transform(new IfFalseNode(parse_predicate));
4614   {
4615     PreserveJVMState pjvms(this);
4616     set_control(if_false);
4617     inc_sp(nargs);
4618     uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
4619   }
4620   Node* if_true = _gvn.transform(new IfTrueNode(parse_predicate));
4621   set_control(if_true);
4622 }
4623 
4624 // Add Parse Predicates which serve as placeholders to create new Runtime Predicates above them. All
4625 // Runtime Predicates inside a Runtime Predicate block share the same uncommon trap as the Parse Predicate.
4626 void GraphKit::add_parse_predicates(int nargs) {
4627   if (ShortRunningLongLoop) {
4628     // Will narrow the limit down with a cast node. Predicates added later may depend on the cast so should be last when
4629     // walking up from the loop.
4630     add_parse_predicate(Deoptimization::Reason_short_running_long_loop, nargs);
4631   }
4632   if (UseLoopPredicate) {
4633     add_parse_predicate(Deoptimization::Reason_predicate, nargs);
4634     if (UseProfiledLoopPredicate) {
4635       add_parse_predicate(Deoptimization::Reason_profile_predicate, nargs);
4636     }
4637   }
4638   if (UseAutoVectorizationPredicate) {
4639     add_parse_predicate(Deoptimization::Reason_auto_vectorization_check, nargs);
4640   }
4641   // Loop Limit Check Predicate should be near the loop.
4642   add_parse_predicate(Deoptimization::Reason_loop_limit_check, nargs);
4643 }
4644 
4645 void GraphKit::sync_kit(IdealKit& ideal) {
4646   reset_memory();
4647   set_all_memory(ideal.merged_memory());
4648   set_i_o(ideal.i_o());
4649   set_control(ideal.ctrl());
4650 }
4651 
4652 void GraphKit::final_sync(IdealKit& ideal) {
4653   // Final sync IdealKit and graphKit.
4654   sync_kit(ideal);
4655 }
4656 
4657 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
4658   Node* len = load_array_length(load_String_value(str, set_ctrl));
4659   Node* coder = load_String_coder(str, set_ctrl);
4660   // Divide length by 2 if coder is UTF16
4661   return _gvn.transform(new RShiftINode(len, coder));
4662 }
4663 
4664 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
4665   int value_offset = java_lang_String::value_offset();
4666   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4667                                                      false, nullptr, Type::Offset(0));
4668   const TypePtr* value_field_type = string_type->add_offset(value_offset);
4669   const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4670                                                   TypeAry::make(TypeInt::BYTE, TypeInt::POS, false, false, true, true, true),
4671                                                   ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0));
4672   Node* p = basic_plus_adr(str, str, value_offset);
4673   Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4674                               IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4675   return load;
4676 }
4677 
4678 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4679   if (!CompactStrings) {
4680     return intcon(java_lang_String::CODER_UTF16);
4681   }
4682   int coder_offset = java_lang_String::coder_offset();
4683   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4684                                                      false, nullptr, Type::Offset(0));
4685   const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4686 
4687   Node* p = basic_plus_adr(str, str, coder_offset);
4688   Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4689                               IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4690   return load;
4691 }
4692 
4693 void GraphKit::store_String_value(Node* str, Node* value) {
4694   int value_offset = java_lang_String::value_offset();
4695   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4696                                                      false, nullptr, Type::Offset(0));
4697   const TypePtr* value_field_type = string_type->add_offset(value_offset);
4698 
4699   access_store_at(str,  basic_plus_adr(str, value_offset), value_field_type,
4700                   value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4701 }
4702 
4703 void GraphKit::store_String_coder(Node* str, Node* value) {
4704   int coder_offset = java_lang_String::coder_offset();
4705   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4706                                                      false, nullptr, Type::Offset(0));
4707   const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4708 
4709   access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4710                   value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4711 }
4712 
4713 // Capture src and dst memory state with a MergeMemNode
4714 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4715   if (src_type == dst_type) {
4716     // Types are equal, we don't need a MergeMemNode
4717     return memory(src_type);
4718   }
4719   MergeMemNode* merge = MergeMemNode::make(map()->memory());
4720   record_for_igvn(merge); // fold it up later, if possible
4721   int src_idx = C->get_alias_index(src_type);
4722   int dst_idx = C->get_alias_index(dst_type);
4723   merge->set_memory_at(src_idx, memory(src_idx));
4724   merge->set_memory_at(dst_idx, memory(dst_idx));
4725   return merge;
4726 }
4727 
4728 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) {
4729   assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported");
4730   assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type");
4731   // If input and output memory types differ, capture both states to preserve
4732   // the dependency between preceding and subsequent loads/stores.
4733   // For example, the following program:
4734   //  StoreB
4735   //  compress_string
4736   //  LoadB
4737   // has this memory graph (use->def):
4738   //  LoadB -> compress_string -> CharMem
4739   //             ... -> StoreB -> ByteMem
4740   // The intrinsic hides the dependency between LoadB and StoreB, causing
4741   // the load to read from memory not containing the result of the StoreB.
4742   // The correct memory graph should look like this:
4743   //  LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem))
4744   Node* mem = capture_memory(src_type, TypeAryPtr::BYTES);
4745   StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count);
4746   Node* res_mem = _gvn.transform(new SCMemProjNode(_gvn.transform(str)));
4747   set_memory(res_mem, TypeAryPtr::BYTES);
4748   return str;
4749 }
4750 
4751 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) {
4752   assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported");
4753   assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type");
4754   // Capture src and dst memory (see comment in 'compress_string').
4755   Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type);
4756   StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count);
4757   set_memory(_gvn.transform(str), dst_type);
4758 }
4759 
4760 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) {
4761   /**
4762    * int i_char = start;
4763    * for (int i_byte = 0; i_byte < count; i_byte++) {
4764    *   dst[i_char++] = (char)(src[i_byte] & 0xff);
4765    * }
4766    */
4767   add_parse_predicates();
4768   C->set_has_loops(true);
4769 
4770   RegionNode* head = new RegionNode(3);
4771   head->init_req(1, control());
4772   gvn().set_type(head, Type::CONTROL);
4773   record_for_igvn(head);
4774 
4775   Node* i_byte = new PhiNode(head, TypeInt::INT);
4776   i_byte->init_req(1, intcon(0));
4777   gvn().set_type(i_byte, TypeInt::INT);
4778   record_for_igvn(i_byte);
4779 
4780   Node* i_char = new PhiNode(head, TypeInt::INT);
4781   i_char->init_req(1, start);
4782   gvn().set_type(i_char, TypeInt::INT);
4783   record_for_igvn(i_char);
4784 
4785   Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES);
4786   gvn().set_type(mem, Type::MEMORY);
4787   record_for_igvn(mem);
4788   set_control(head);
4789   set_memory(mem, TypeAryPtr::BYTES);
4790   Node* ch = load_array_element(src, i_byte, TypeAryPtr::BYTES, /* set_ctrl */ true);
4791   Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE),
4792                              AndI(ch, intcon(0xff)), T_CHAR, MemNode::unordered, false,
4793                              false, true /* mismatched */);
4794 
4795   IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
4796   head->init_req(2, IfTrue(iff));
4797   mem->init_req(2, st);
4798   i_byte->init_req(2, AddI(i_byte, intcon(1)));
4799   i_char->init_req(2, AddI(i_char, intcon(2)));
4800 
4801   set_control(IfFalse(iff));
4802   set_memory(st, TypeAryPtr::BYTES);
4803 }
4804 
4805 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4806   if (!field->is_constant()) {
4807     return nullptr; // Field not marked as constant.
4808   }
4809   ciInstance* holder = nullptr;
4810   if (!field->is_static()) {
4811     ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4812     if (const_oop != nullptr && const_oop->is_instance()) {
4813       holder = const_oop->as_instance();
4814     }
4815   }
4816   const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4817                                                         /*is_unsigned_load=*/false);
4818   if (con_type != nullptr) {
4819     Node* con = makecon(con_type);
4820     if (field->type()->is_inlinetype()) {
4821       con = InlineTypeNode::make_from_oop(this, con, field->type()->as_inline_klass());
4822     } else if (con_type->is_inlinetypeptr()) {
4823       con = InlineTypeNode::make_from_oop(this, con, con_type->inline_klass());
4824     }
4825     return con;
4826   }
4827   return nullptr;
4828 }
4829 
4830 Node* GraphKit::maybe_narrow_object_type(Node* obj, ciKlass* type) {
4831   const Type* obj_type = obj->bottom_type();
4832   const TypeOopPtr* sig_type = TypeOopPtr::make_from_klass(type);
4833   if (obj_type->isa_oopptr() && sig_type->is_loaded() && !obj_type->higher_equal(sig_type)) {
4834     const Type* narrow_obj_type = obj_type->filter_speculative(sig_type); // keep speculative part
4835     Node* casted_obj = gvn().transform(new CheckCastPPNode(control(), obj, narrow_obj_type));
4836     obj = casted_obj;
4837   }
4838   if (sig_type->is_inlinetypeptr()) {
4839     obj = InlineTypeNode::make_from_oop(this, obj, sig_type->inline_klass());
4840   }
4841   return obj;
4842 }