1 /*
   2  * Copyright (c) 2018, 2019, Red Hat, Inc. All rights reserved.
   3  *
   4  * This code is free software; you can redistribute it and/or modify it
   5  * under the terms of the GNU General Public License version 2 only, as
   6  * published by the Free Software Foundation.
   7  *
   8  * This code is distributed in the hope that it will be useful, but WITHOUT
   9  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  10  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  11  * version 2 for more details (a copy is included in the LICENSE file that
  12  * accompanied this code).
  13  *
  14  * You should have received a copy of the GNU General Public License version
  15  * 2 along with this work; if not, write to the Free Software Foundation,
  16  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  17  *
  18  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  19  * or visit www.oracle.com if you need additional information or have any
  20  * questions.
  21  *
  22  */
  23 
  24 #include "precompiled.hpp"
  25 #include "gc/shared/barrierSet.hpp"
  26 #include "gc/shenandoah/shenandoahForwarding.hpp"
  27 #include "gc/shenandoah/shenandoahHeap.hpp"
  28 #include "gc/shenandoah/shenandoahHeuristics.hpp"
  29 #include "gc/shenandoah/shenandoahRuntime.hpp"
  30 #include "gc/shenandoah/shenandoahThreadLocalData.hpp"
  31 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
  32 #include "gc/shenandoah/c2/shenandoahSupport.hpp"
  33 #include "opto/arraycopynode.hpp"
  34 #include "opto/escape.hpp"
  35 #include "opto/graphKit.hpp"
  36 #include "opto/idealKit.hpp"
  37 #include "opto/macro.hpp"
  38 #include "opto/movenode.hpp"
  39 #include "opto/narrowptrnode.hpp"
  40 #include "opto/rootnode.hpp"
  41 
  42 ShenandoahBarrierSetC2* ShenandoahBarrierSetC2::bsc2() {
  43   return reinterpret_cast<ShenandoahBarrierSetC2*>(BarrierSet::barrier_set()->barrier_set_c2());
  44 }
  45 
  46 ShenandoahBarrierSetC2State::ShenandoahBarrierSetC2State(Arena* comp_arena)
  47   : _enqueue_barriers(new (comp_arena) GrowableArray<ShenandoahEnqueueBarrierNode*>(comp_arena, 8,  0, NULL)),
  48     _load_reference_barriers(new (comp_arena) GrowableArray<ShenandoahLoadReferenceBarrierNode*>(comp_arena, 8,  0, NULL)) {
  49 }
  50 
  51 int ShenandoahBarrierSetC2State::enqueue_barriers_count() const {
  52   return _enqueue_barriers->length();
  53 }
  54 
  55 ShenandoahEnqueueBarrierNode* ShenandoahBarrierSetC2State::enqueue_barrier(int idx) const {
  56   return _enqueue_barriers->at(idx);
  57 }
  58 
  59 void ShenandoahBarrierSetC2State::add_enqueue_barrier(ShenandoahEnqueueBarrierNode * n) {
  60   assert(!_enqueue_barriers->contains(n), "duplicate entry in barrier list");
  61   _enqueue_barriers->append(n);
  62 }
  63 
  64 void ShenandoahBarrierSetC2State::remove_enqueue_barrier(ShenandoahEnqueueBarrierNode * n) {
  65   if (_enqueue_barriers->contains(n)) {
  66     _enqueue_barriers->remove(n);
  67   }
  68 }
  69 
  70 int ShenandoahBarrierSetC2State::load_reference_barriers_count() const {
  71   return _load_reference_barriers->length();
  72 }
  73 
  74 ShenandoahLoadReferenceBarrierNode* ShenandoahBarrierSetC2State::load_reference_barrier(int idx) const {
  75   return _load_reference_barriers->at(idx);
  76 }
  77 
  78 void ShenandoahBarrierSetC2State::add_load_reference_barrier(ShenandoahLoadReferenceBarrierNode * n) {
  79   assert(!_load_reference_barriers->contains(n), "duplicate entry in barrier list");
  80   _load_reference_barriers->append(n);
  81 }
  82 
  83 void ShenandoahBarrierSetC2State::remove_load_reference_barrier(ShenandoahLoadReferenceBarrierNode * n) {
  84   if (_load_reference_barriers->contains(n)) {
  85     _load_reference_barriers->remove(n);
  86   }
  87 }
  88 
  89 Node* ShenandoahBarrierSetC2::shenandoah_storeval_barrier(GraphKit* kit, Node* obj) const {
  90   if (ShenandoahStoreValEnqueueBarrier) {
  91     obj = shenandoah_enqueue_barrier(kit, obj);
  92   }
  93   return obj;
  94 }
  95 
  96 #define __ kit->
  97 
  98 bool ShenandoahBarrierSetC2::satb_can_remove_pre_barrier(GraphKit* kit, PhaseTransform* phase, Node* adr,
  99                                                          BasicType bt, uint adr_idx) const {
 100   intptr_t offset = 0;
 101   Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
 102   AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
 103 
 104   if (offset == Type::OffsetBot) {
 105     return false; // cannot unalias unless there are precise offsets
 106   }
 107 
 108   if (alloc == NULL) {
 109     return false; // No allocation found
 110   }
 111 
 112   intptr_t size_in_bytes = type2aelembytes(bt);
 113 
 114   Node* mem = __ memory(adr_idx); // start searching here...
 115 
 116   for (int cnt = 0; cnt < 50; cnt++) {
 117 
 118     if (mem->is_Store()) {
 119 
 120       Node* st_adr = mem->in(MemNode::Address);
 121       intptr_t st_offset = 0;
 122       Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
 123 
 124       if (st_base == NULL) {
 125         break; // inscrutable pointer
 126       }
 127 
 128       // Break we have found a store with same base and offset as ours so break
 129       if (st_base == base && st_offset == offset) {
 130         break;
 131       }
 132 
 133       if (st_offset != offset && st_offset != Type::OffsetBot) {
 134         const int MAX_STORE = BytesPerLong;
 135         if (st_offset >= offset + size_in_bytes ||
 136             st_offset <= offset - MAX_STORE ||
 137             st_offset <= offset - mem->as_Store()->memory_size()) {
 138           // Success:  The offsets are provably independent.
 139           // (You may ask, why not just test st_offset != offset and be done?
 140           // The answer is that stores of different sizes can co-exist
 141           // in the same sequence of RawMem effects.  We sometimes initialize
 142           // a whole 'tile' of array elements with a single jint or jlong.)
 143           mem = mem->in(MemNode::Memory);
 144           continue; // advance through independent store memory
 145         }
 146       }
 147 
 148       if (st_base != base
 149           && MemNode::detect_ptr_independence(base, alloc, st_base,
 150                                               AllocateNode::Ideal_allocation(st_base, phase),
 151                                               phase)) {
 152         // Success:  The bases are provably independent.
 153         mem = mem->in(MemNode::Memory);
 154         continue; // advance through independent store memory
 155       }
 156     } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
 157 
 158       InitializeNode* st_init = mem->in(0)->as_Initialize();
 159       AllocateNode* st_alloc = st_init->allocation();
 160 
 161       // Make sure that we are looking at the same allocation site.
 162       // The alloc variable is guaranteed to not be null here from earlier check.
 163       if (alloc == st_alloc) {
 164         // Check that the initialization is storing NULL so that no previous store
 165         // has been moved up and directly write a reference
 166         Node* captured_store = st_init->find_captured_store(offset,
 167                                                             type2aelembytes(T_OBJECT),
 168                                                             phase);
 169         if (captured_store == NULL || captured_store == st_init->zero_memory()) {
 170           return true;
 171         }
 172       }
 173     }
 174 
 175     // Unless there is an explicit 'continue', we must bail out here,
 176     // because 'mem' is an inscrutable memory state (e.g., a call).
 177     break;
 178   }
 179 
 180   return false;
 181 }
 182 
 183 #undef __
 184 #define __ ideal.
 185 
 186 void ShenandoahBarrierSetC2::satb_write_barrier_pre(GraphKit* kit,
 187                                                     bool do_load,
 188                                                     Node* obj,
 189                                                     Node* adr,
 190                                                     uint alias_idx,
 191                                                     Node* val,
 192                                                     const TypeOopPtr* val_type,
 193                                                     Node* pre_val,
 194                                                     BasicType bt) const {
 195   // Some sanity checks
 196   // Note: val is unused in this routine.
 197 
 198   if (do_load) {
 199     // We need to generate the load of the previous value
 200     assert(obj != NULL, "must have a base");
 201     assert(adr != NULL, "where are loading from?");
 202     assert(pre_val == NULL, "loaded already?");
 203     assert(val_type != NULL, "need a type");
 204 
 205     if (ReduceInitialCardMarks
 206         && satb_can_remove_pre_barrier(kit, &kit->gvn(), adr, bt, alias_idx)) {
 207       return;
 208     }
 209 
 210   } else {
 211     // In this case both val_type and alias_idx are unused.
 212     assert(pre_val != NULL, "must be loaded already");
 213     // Nothing to be done if pre_val is null.
 214     if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
 215     assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
 216   }
 217   assert(bt == T_OBJECT, "or we shouldn't be here");
 218 
 219   IdealKit ideal(kit, true);
 220 
 221   Node* tls = __ thread(); // ThreadLocalStorage
 222 
 223   Node* no_base = __ top();
 224   Node* zero  = __ ConI(0);
 225   Node* zeroX = __ ConX(0);
 226 
 227   float likely  = PROB_LIKELY(0.999);
 228   float unlikely  = PROB_UNLIKELY(0.999);
 229 
 230   // Offsets into the thread
 231   const int index_offset   = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset());
 232   const int buffer_offset  = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
 233 
 234   // Now the actual pointers into the thread
 235   Node* buffer_adr  = __ AddP(no_base, tls, __ ConX(buffer_offset));
 236   Node* index_adr   = __ AddP(no_base, tls, __ ConX(index_offset));
 237 
 238   // Now some of the values
 239   Node* marking;
 240   Node* gc_state = __ AddP(no_base, tls, __ ConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset())));
 241   Node* ld = __ load(__ ctrl(), gc_state, TypeInt::BYTE, T_BYTE, Compile::AliasIdxRaw);
 242   marking = __ AndI(ld, __ ConI(ShenandoahHeap::MARKING));
 243   assert(ShenandoahBarrierC2Support::is_gc_state_load(ld), "Should match the shape");
 244 
 245   // if (!marking)
 246   __ if_then(marking, BoolTest::ne, zero, unlikely); {
 247     BasicType index_bt = TypeX_X->basic_type();
 248     assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size.");
 249     Node* index   = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
 250 
 251     if (do_load) {
 252       // load original value
 253       // alias_idx correct??
 254       pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
 255     }
 256 
 257     // if (pre_val != NULL)
 258     __ if_then(pre_val, BoolTest::ne, kit->null()); {
 259       Node* buffer  = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
 260 
 261       // is the queue for this thread full?
 262       __ if_then(index, BoolTest::ne, zeroX, likely); {
 263 
 264         // decrement the index
 265         Node* next_index = kit->gvn().transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
 266 
 267         // Now get the buffer location we will log the previous value into and store it
 268         Node *log_addr = __ AddP(no_base, buffer, next_index);
 269         __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
 270         // update the index
 271         __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
 272 
 273       } __ else_(); {
 274 
 275         // logging buffer is full, call the runtime
 276         const TypeFunc *tf = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type();
 277         __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), "shenandoah_wb_pre", pre_val, tls);
 278       } __ end_if();  // (!index)
 279     } __ end_if();  // (pre_val != NULL)
 280   } __ end_if();  // (!marking)
 281 
 282   // Final sync IdealKit and GraphKit.
 283   kit->final_sync(ideal);
 284 
 285   if (ShenandoahSATBBarrier && adr != NULL) {
 286     Node* c = kit->control();
 287     Node* call = c->in(1)->in(1)->in(1)->in(0);
 288     assert(is_shenandoah_wb_pre_call(call), "shenandoah_wb_pre call expected");
 289     call->add_req(adr);
 290   }
 291 }
 292 
 293 bool ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(Node* call) {
 294   return call->is_CallLeaf() &&
 295          call->as_CallLeaf()->entry_point() == CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry);
 296 }
 297 
 298 bool ShenandoahBarrierSetC2::is_shenandoah_lrb_call(Node* call) {
 299   return call->is_CallLeaf() &&
 300           call->as_CallLeaf()->entry_point() == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier);
 301 }
 302 
 303 bool ShenandoahBarrierSetC2::is_shenandoah_marking_if(PhaseTransform *phase, Node* n) {
 304   if (n->Opcode() != Op_If) {
 305     return false;
 306   }
 307 
 308   Node* bol = n->in(1);
 309   assert(bol->is_Bool(), "");
 310   Node* cmpx = bol->in(1);
 311   if (bol->as_Bool()->_test._test == BoolTest::ne &&
 312       cmpx->is_Cmp() && cmpx->in(2) == phase->intcon(0) &&
 313       is_shenandoah_state_load(cmpx->in(1)->in(1)) &&
 314       cmpx->in(1)->in(2)->is_Con() &&
 315       cmpx->in(1)->in(2) == phase->intcon(ShenandoahHeap::MARKING)) {
 316     return true;
 317   }
 318 
 319   return false;
 320 }
 321 
 322 bool ShenandoahBarrierSetC2::is_shenandoah_state_load(Node* n) {
 323   if (!n->is_Load()) return false;
 324   const int state_offset = in_bytes(ShenandoahThreadLocalData::gc_state_offset());
 325   return n->in(2)->is_AddP() && n->in(2)->in(2)->Opcode() == Op_ThreadLocal
 326          && n->in(2)->in(3)->is_Con()
 327          && n->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == state_offset;
 328 }
 329 
 330 void ShenandoahBarrierSetC2::shenandoah_write_barrier_pre(GraphKit* kit,
 331                                                           bool do_load,
 332                                                           Node* obj,
 333                                                           Node* adr,
 334                                                           uint alias_idx,
 335                                                           Node* val,
 336                                                           const TypeOopPtr* val_type,
 337                                                           Node* pre_val,
 338                                                           BasicType bt) const {
 339   if (ShenandoahSATBBarrier) {
 340     IdealKit ideal(kit);
 341     kit->sync_kit(ideal);
 342 
 343     satb_write_barrier_pre(kit, do_load, obj, adr, alias_idx, val, val_type, pre_val, bt);
 344 
 345     ideal.sync_kit(kit);
 346     kit->final_sync(ideal);
 347   }
 348 }
 349 
 350 Node* ShenandoahBarrierSetC2::shenandoah_enqueue_barrier(GraphKit* kit, Node* pre_val) const {
 351   return kit->gvn().transform(new ShenandoahEnqueueBarrierNode(pre_val));
 352 }
 353 
 354 // Helper that guards and inserts a pre-barrier.
 355 void ShenandoahBarrierSetC2::insert_pre_barrier(GraphKit* kit, Node* base_oop, Node* offset,
 356                                                 Node* pre_val, bool need_mem_bar) const {
 357   // We could be accessing the referent field of a reference object. If so, when G1
 358   // is enabled, we need to log the value in the referent field in an SATB buffer.
 359   // This routine performs some compile time filters and generates suitable
 360   // runtime filters that guard the pre-barrier code.
 361   // Also add memory barrier for non volatile load from the referent field
 362   // to prevent commoning of loads across safepoint.
 363 
 364   // Some compile time checks.
 365 
 366   // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
 367   const TypeX* otype = offset->find_intptr_t_type();
 368   if (otype != NULL && otype->is_con() &&
 369       otype->get_con() != java_lang_ref_Reference::referent_offset) {
 370     // Constant offset but not the reference_offset so just return
 371     return;
 372   }
 373 
 374   // We only need to generate the runtime guards for instances.
 375   const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
 376   if (btype != NULL) {
 377     if (btype->isa_aryptr()) {
 378       // Array type so nothing to do
 379       return;
 380     }
 381 
 382     const TypeInstPtr* itype = btype->isa_instptr();
 383     if (itype != NULL) {
 384       // Can the klass of base_oop be statically determined to be
 385       // _not_ a sub-class of Reference and _not_ Object?
 386       ciKlass* klass = itype->klass();
 387       if ( klass->is_loaded() &&
 388           !klass->is_subtype_of(kit->env()->Reference_klass()) &&
 389           !kit->env()->Object_klass()->is_subtype_of(klass)) {
 390         return;
 391       }
 392     }
 393   }
 394 
 395   // The compile time filters did not reject base_oop/offset so
 396   // we need to generate the following runtime filters
 397   //
 398   // if (offset == java_lang_ref_Reference::_reference_offset) {
 399   //   if (instance_of(base, java.lang.ref.Reference)) {
 400   //     pre_barrier(_, pre_val, ...);
 401   //   }
 402   // }
 403 
 404   float likely   = PROB_LIKELY(  0.999);
 405   float unlikely = PROB_UNLIKELY(0.999);
 406 
 407   IdealKit ideal(kit);
 408 
 409   Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
 410 
 411   __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
 412       // Update graphKit memory and control from IdealKit.
 413       kit->sync_kit(ideal);
 414 
 415       Node* ref_klass_con = kit->makecon(TypeKlassPtr::make(kit->env()->Reference_klass()));
 416       Node* is_instof = kit->gen_instanceof(base_oop, ref_klass_con);
 417 
 418       // Update IdealKit memory and control from graphKit.
 419       __ sync_kit(kit);
 420 
 421       Node* one = __ ConI(1);
 422       // is_instof == 0 if base_oop == NULL
 423       __ if_then(is_instof, BoolTest::eq, one, unlikely); {
 424 
 425         // Update graphKit from IdeakKit.
 426         kit->sync_kit(ideal);
 427 
 428         // Use the pre-barrier to record the value in the referent field
 429         satb_write_barrier_pre(kit, false /* do_load */,
 430                                NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
 431                                pre_val /* pre_val */,
 432                                T_OBJECT);
 433         if (need_mem_bar) {
 434           // Add memory barrier to prevent commoning reads from this field
 435           // across safepoint since GC can change its value.
 436           kit->insert_mem_bar(Op_MemBarCPUOrder);
 437         }
 438         // Update IdealKit from graphKit.
 439         __ sync_kit(kit);
 440 
 441       } __ end_if(); // _ref_type != ref_none
 442   } __ end_if(); // offset == referent_offset
 443 
 444   // Final sync IdealKit and GraphKit.
 445   kit->final_sync(ideal);
 446 }
 447 
 448 #undef __
 449 
 450 const TypeFunc* ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type() {
 451   const Type **fields = TypeTuple::fields(2);
 452   fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value
 453   fields[TypeFunc::Parms+1] = TypeRawPtr::NOTNULL; // thread
 454   const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields);
 455 
 456   // create result type (range)
 457   fields = TypeTuple::fields(0);
 458   const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields);
 459 
 460   return TypeFunc::make(domain, range);
 461 }
 462 
 463 const TypeFunc* ShenandoahBarrierSetC2::shenandoah_clone_barrier_Type() {
 464   const Type **fields = TypeTuple::fields(3);
 465   fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // src
 466   fields[TypeFunc::Parms+1] = TypeInstPtr::NOTNULL; // dst
 467   fields[TypeFunc::Parms+2] = TypeInt::INT; // length
 468   const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+3, fields);
 469 
 470   // create result type (range)
 471   fields = TypeTuple::fields(0);
 472   const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields);
 473 
 474   return TypeFunc::make(domain, range);
 475 }
 476 
 477 const TypeFunc* ShenandoahBarrierSetC2::shenandoah_load_reference_barrier_Type() {
 478   const Type **fields = TypeTuple::fields(1);
 479   fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value
 480   const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+1, fields);
 481 
 482   // create result type (range)
 483   fields = TypeTuple::fields(1);
 484   fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL;
 485   const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+1, fields);
 486 
 487   return TypeFunc::make(domain, range);
 488 }
 489 
 490 Node* ShenandoahBarrierSetC2::store_at_resolved(C2Access& access, C2AccessValue& val) const {
 491   DecoratorSet decorators = access.decorators();
 492 
 493   const TypePtr* adr_type = access.addr().type();
 494   Node* adr = access.addr().node();
 495 
 496   bool anonymous = (decorators & ON_UNKNOWN_OOP_REF) != 0;
 497   bool on_heap = (decorators & IN_HEAP) != 0;
 498 
 499   if (!access.is_oop() || (!on_heap && !anonymous)) {
 500     return BarrierSetC2::store_at_resolved(access, val);
 501   }
 502 
 503   if (access.is_parse_access()) {
 504     C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access);
 505     GraphKit* kit = parse_access.kit();
 506 
 507     uint adr_idx = kit->C->get_alias_index(adr_type);
 508     assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
 509     Node* value = val.node();
 510     value = shenandoah_storeval_barrier(kit, value);
 511     val.set_node(value);
 512     shenandoah_write_barrier_pre(kit, true /* do_load */, /*kit->control(),*/ access.base(), adr, adr_idx, val.node(),
 513                                  static_cast<const TypeOopPtr*>(val.type()), NULL /* pre_val */, access.type());
 514   } else {
 515     assert(access.is_opt_access(), "only for optimization passes");
 516     assert(((decorators & C2_TIGHTLY_COUPLED_ALLOC) != 0 || !ShenandoahSATBBarrier) && (decorators & C2_ARRAY_COPY) != 0, "unexpected caller of this code");
 517     C2OptAccess& opt_access = static_cast<C2OptAccess&>(access);
 518     PhaseGVN& gvn =  opt_access.gvn();
 519     MergeMemNode* mm = opt_access.mem();
 520 
 521     if (ShenandoahStoreValEnqueueBarrier) {
 522       Node* enqueue = gvn.transform(new ShenandoahEnqueueBarrierNode(val.node()));
 523       val.set_node(enqueue);
 524     }
 525   }
 526   return BarrierSetC2::store_at_resolved(access, val);
 527 }
 528 
 529 Node* ShenandoahBarrierSetC2::load_at_resolved(C2Access& access, const Type* val_type) const {
 530   DecoratorSet decorators = access.decorators();
 531 
 532   Node* adr = access.addr().node();
 533   Node* obj = access.base();
 534 
 535   bool mismatched = (decorators & C2_MISMATCHED) != 0;
 536   bool unknown = (decorators & ON_UNKNOWN_OOP_REF) != 0;
 537   bool on_heap = (decorators & IN_HEAP) != 0;
 538   bool on_weak = (decorators & ON_WEAK_OOP_REF) != 0;
 539   bool is_unordered = (decorators & MO_UNORDERED) != 0;
 540   bool need_cpu_mem_bar = !is_unordered || mismatched || !on_heap;
 541 
 542   Node* top = Compile::current()->top();
 543 
 544   Node* offset = adr->is_AddP() ? adr->in(AddPNode::Offset) : top;
 545   Node* load = BarrierSetC2::load_at_resolved(access, val_type);
 546 
 547   if (access.is_oop()) {
 548     if (ShenandoahLoadRefBarrier) {
 549       load = new ShenandoahLoadReferenceBarrierNode(NULL, load, (decorators & IN_NATIVE) != 0);
 550       if (access.is_parse_access()) {
 551         load = static_cast<C2ParseAccess &>(access).kit()->gvn().transform(load);
 552       } else {
 553         load = static_cast<C2OptAccess &>(access).gvn().transform(load);
 554       }
 555     }
 556   }
 557 
 558   // If we are reading the value of the referent field of a Reference
 559   // object (either by using Unsafe directly or through reflection)
 560   // then, if SATB is enabled, we need to record the referent in an
 561   // SATB log buffer using the pre-barrier mechanism.
 562   // Also we need to add memory barrier to prevent commoning reads
 563   // from this field across safepoint since GC can change its value.
 564   bool need_read_barrier = ShenandoahKeepAliveBarrier &&
 565     (on_heap && (on_weak || (unknown && offset != top && obj != top)));
 566 
 567   if (!access.is_oop() || !need_read_barrier) {
 568     return load;
 569   }
 570 
 571   assert(access.is_parse_access(), "entry not supported at optimization time");
 572   C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access);
 573   GraphKit* kit = parse_access.kit();
 574 
 575   if (on_weak) {
 576     // Use the pre-barrier to record the value in the referent field
 577     satb_write_barrier_pre(kit, false /* do_load */,
 578                            NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
 579                            load /* pre_val */, T_OBJECT);
 580     // Add memory barrier to prevent commoning reads from this field
 581     // across safepoint since GC can change its value.
 582     kit->insert_mem_bar(Op_MemBarCPUOrder);
 583   } else if (unknown) {
 584     // We do not require a mem bar inside pre_barrier if need_mem_bar
 585     // is set: the barriers would be emitted by us.
 586     insert_pre_barrier(kit, obj, offset, load, !need_cpu_mem_bar);
 587   }
 588 
 589   return load;
 590 }
 591 
 592 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_val_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
 593                                                    Node* new_val, const Type* value_type) const {
 594   GraphKit* kit = access.kit();
 595   if (access.is_oop()) {
 596     new_val = shenandoah_storeval_barrier(kit, new_val);
 597     shenandoah_write_barrier_pre(kit, false /* do_load */,
 598                                  NULL, NULL, max_juint, NULL, NULL,
 599                                  expected_val /* pre_val */, T_OBJECT);
 600 
 601     MemNode::MemOrd mo = access.mem_node_mo();
 602     Node* mem = access.memory();
 603     Node* adr = access.addr().node();
 604     const TypePtr* adr_type = access.addr().type();
 605     Node* load_store = NULL;
 606 
 607 #ifdef _LP64
 608     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
 609       Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop()));
 610       Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop()));
 611       if (ShenandoahCASBarrier) {
 612         load_store = kit->gvn().transform(new ShenandoahCompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
 613       } else {
 614         load_store = kit->gvn().transform(new CompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
 615       }
 616     } else
 617 #endif
 618     {
 619       if (ShenandoahCASBarrier) {
 620         load_store = kit->gvn().transform(new ShenandoahCompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo));
 621       } else {
 622         load_store = kit->gvn().transform(new CompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo));
 623       }
 624     }
 625 
 626     access.set_raw_access(load_store);
 627     pin_atomic_op(access);
 628 
 629 #ifdef _LP64
 630     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
 631       load_store = kit->gvn().transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
 632     }
 633 #endif
 634     load_store = kit->gvn().transform(new ShenandoahLoadReferenceBarrierNode(NULL, load_store, false));
 635     return load_store;
 636   }
 637   return BarrierSetC2::atomic_cmpxchg_val_at_resolved(access, expected_val, new_val, value_type);
 638 }
 639 
 640 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_bool_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
 641                                                               Node* new_val, const Type* value_type) const {
 642   GraphKit* kit = access.kit();
 643   if (access.is_oop()) {
 644     new_val = shenandoah_storeval_barrier(kit, new_val);
 645     shenandoah_write_barrier_pre(kit, false /* do_load */,
 646                                  NULL, NULL, max_juint, NULL, NULL,
 647                                  expected_val /* pre_val */, T_OBJECT);
 648     DecoratorSet decorators = access.decorators();
 649     MemNode::MemOrd mo = access.mem_node_mo();
 650     Node* mem = access.memory();
 651     bool is_weak_cas = (decorators & C2_WEAK_CMPXCHG) != 0;
 652     Node* load_store = NULL;
 653     Node* adr = access.addr().node();
 654 #ifdef _LP64
 655     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
 656       Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop()));
 657       Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop()));
 658       if (ShenandoahCASBarrier) {
 659         if (is_weak_cas) {
 660           load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
 661         } else {
 662           load_store = kit->gvn().transform(new ShenandoahCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
 663         }
 664       } else {
 665         if (is_weak_cas) {
 666           load_store = kit->gvn().transform(new WeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
 667         } else {
 668           load_store = kit->gvn().transform(new CompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
 669         }
 670       }
 671     } else
 672 #endif
 673     {
 674       if (ShenandoahCASBarrier) {
 675         if (is_weak_cas) {
 676           load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
 677         } else {
 678           load_store = kit->gvn().transform(new ShenandoahCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
 679         }
 680       } else {
 681         if (is_weak_cas) {
 682           load_store = kit->gvn().transform(new WeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
 683         } else {
 684           load_store = kit->gvn().transform(new CompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
 685         }
 686       }
 687     }
 688     access.set_raw_access(load_store);
 689     pin_atomic_op(access);
 690     return load_store;
 691   }
 692   return BarrierSetC2::atomic_cmpxchg_bool_at_resolved(access, expected_val, new_val, value_type);
 693 }
 694 
 695 Node* ShenandoahBarrierSetC2::atomic_xchg_at_resolved(C2AtomicParseAccess& access, Node* val, const Type* value_type) const {
 696   GraphKit* kit = access.kit();
 697   if (access.is_oop()) {
 698     val = shenandoah_storeval_barrier(kit, val);
 699   }
 700   Node* result = BarrierSetC2::atomic_xchg_at_resolved(access, val, value_type);
 701   if (access.is_oop()) {
 702     result = kit->gvn().transform(new ShenandoahLoadReferenceBarrierNode(NULL, result, false));
 703     shenandoah_write_barrier_pre(kit, false /* do_load */,
 704                                  NULL, NULL, max_juint, NULL, NULL,
 705                                  result /* pre_val */, T_OBJECT);
 706   }
 707   return result;
 708 }
 709 
 710 // Support for GC barriers emitted during parsing
 711 bool ShenandoahBarrierSetC2::is_gc_barrier_node(Node* node) const {
 712   if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) return true;
 713   if (node->Opcode() != Op_CallLeaf && node->Opcode() != Op_CallLeafNoFP) {
 714     return false;
 715   }
 716   CallLeafNode *call = node->as_CallLeaf();
 717   if (call->_name == NULL) {
 718     return false;
 719   }
 720 
 721   return strcmp(call->_name, "shenandoah_clone_barrier") == 0 ||
 722          strcmp(call->_name, "shenandoah_cas_obj") == 0 ||
 723          strcmp(call->_name, "shenandoah_wb_pre") == 0;
 724 }
 725 
 726 Node* ShenandoahBarrierSetC2::step_over_gc_barrier(Node* c) const {
 727   if (c->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
 728     return c->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
 729   }
 730   if (c->Opcode() == Op_ShenandoahEnqueueBarrier) {
 731     c = c->in(1);
 732   }
 733   return c;
 734 }
 735 
 736 bool ShenandoahBarrierSetC2::expand_barriers(Compile* C, PhaseIterGVN& igvn) const {
 737   return !ShenandoahBarrierC2Support::expand(C, igvn);
 738 }
 739 
 740 bool ShenandoahBarrierSetC2::optimize_loops(PhaseIdealLoop* phase, LoopOptsMode mode, VectorSet& visited, Node_Stack& nstack, Node_List& worklist) const {
 741   if (mode == LoopOptsShenandoahExpand) {
 742     assert(UseShenandoahGC, "only for shenandoah");
 743     ShenandoahBarrierC2Support::pin_and_expand(phase);
 744     return true;
 745   } else if (mode == LoopOptsShenandoahPostExpand) {
 746     assert(UseShenandoahGC, "only for shenandoah");
 747     visited.Clear();
 748     ShenandoahBarrierC2Support::optimize_after_expansion(visited, nstack, worklist, phase);
 749     return true;
 750   }
 751   return false;
 752 }
 753 
 754 bool ShenandoahBarrierSetC2::array_copy_requires_gc_barriers(bool tightly_coupled_alloc, BasicType type, bool is_clone, ArrayCopyPhase phase) const {
 755   bool is_oop = type == T_OBJECT || type == T_ARRAY;
 756   if (!is_oop) {
 757     return false;
 758   }
 759   if (tightly_coupled_alloc) {
 760     if (phase == Optimization) {
 761       return false;
 762     }
 763     return !is_clone;
 764   }
 765   if (phase == Optimization) {
 766     return !ShenandoahStoreValEnqueueBarrier;
 767   }
 768   return true;
 769 }
 770 
 771 bool ShenandoahBarrierSetC2::clone_needs_barrier(Node* src, PhaseGVN& gvn) {
 772   const TypeOopPtr* src_type = gvn.type(src)->is_oopptr();
 773   if (src_type->isa_instptr() != NULL) {
 774     ciInstanceKlass* ik = src_type->klass()->as_instance_klass();
 775     if ((src_type->klass_is_exact() || (!ik->is_interface() && !ik->has_subklass())) && !ik->has_injected_fields()) {
 776       if (ik->has_object_fields()) {
 777         return true;
 778       } else {
 779         if (!src_type->klass_is_exact()) {
 780           Compile::current()->dependencies()->assert_leaf_type(ik);
 781         }
 782       }
 783     } else {
 784       return true;
 785         }
 786   } else if (src_type->isa_aryptr()) {
 787     BasicType src_elem  = src_type->klass()->as_array_klass()->element_type()->basic_type();
 788     if (src_elem == T_OBJECT || src_elem == T_ARRAY) {
 789       return true;
 790     }
 791   } else {
 792     return true;
 793   }
 794   return false;
 795 }
 796 
 797 #define XTOP LP64_ONLY(COMMA phase->top())
 798 
 799 void ShenandoahBarrierSetC2::clone_at_expansion(PhaseMacroExpand* phase, ArrayCopyNode* ac) const {
 800   Node* ctrl = ac->in(TypeFunc::Control);
 801   Node* mem = ac->in(TypeFunc::Memory);
 802   Node* src = ac->in(ArrayCopyNode::Src);
 803   Node* src_offset = ac->in(ArrayCopyNode::SrcPos);
 804   Node* dest = ac->in(ArrayCopyNode::Dest);
 805   Node* dest_offset = ac->in(ArrayCopyNode::DestPos);
 806   Node* length = ac->in(ArrayCopyNode::Length);
 807   assert (src_offset == NULL && dest_offset == NULL, "for clone offsets should be null");
 808   if (ShenandoahCloneBarrier && clone_needs_barrier(src, phase->igvn())) {
 809     const char* copyfunc_name = "shenandoah_clone";
 810     address     copyfunc_addr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::shenandoah_clone_barrier);
 811     const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
 812     const TypeFunc* call_type = ShenandoahBarrierSetC2::shenandoah_clone_barrier_Type();
 813     Node* call = phase->make_leaf_call(ctrl, mem, call_type, copyfunc_addr, copyfunc_name, raw_adr_type, src, dest, length);
 814     call = phase->transform_later(call);
 815     phase->igvn().replace_node(ac, call);
 816   } else {
 817     BarrierSetC2::clone_at_expansion(phase, ac);
 818   }
 819 }
 820 
 821 
 822 // Support for macro expanded GC barriers
 823 void ShenandoahBarrierSetC2::register_potential_barrier_node(Node* node) const {
 824   if (node->Opcode() == Op_ShenandoahEnqueueBarrier) {
 825     state()->add_enqueue_barrier((ShenandoahEnqueueBarrierNode*) node);
 826   }
 827   if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
 828     state()->add_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node);
 829   }
 830 }
 831 
 832 void ShenandoahBarrierSetC2::unregister_potential_barrier_node(Node* node) const {
 833   if (node->Opcode() == Op_ShenandoahEnqueueBarrier) {
 834     state()->remove_enqueue_barrier((ShenandoahEnqueueBarrierNode*) node);
 835   }
 836   if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
 837     state()->remove_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node);
 838   }
 839 }
 840 
 841 void ShenandoahBarrierSetC2::eliminate_gc_barrier(PhaseMacroExpand* macro, Node* n) const {
 842   if (is_shenandoah_wb_pre_call(n)) {
 843     shenandoah_eliminate_wb_pre(n, &macro->igvn());
 844   }
 845 }
 846 
 847 void ShenandoahBarrierSetC2::shenandoah_eliminate_wb_pre(Node* call, PhaseIterGVN* igvn) const {
 848   assert(UseShenandoahGC && is_shenandoah_wb_pre_call(call), "");
 849   Node* c = call->as_Call()->proj_out(TypeFunc::Control);
 850   c = c->unique_ctrl_out();
 851   assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
 852   c = c->unique_ctrl_out();
 853   assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
 854   Node* iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
 855   assert(iff->is_If(), "expect test");
 856   if (!is_shenandoah_marking_if(igvn, iff)) {
 857     c = c->unique_ctrl_out();
 858     assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
 859     iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
 860     assert(is_shenandoah_marking_if(igvn, iff), "expect marking test");
 861   }
 862   Node* cmpx = iff->in(1)->in(1);
 863   igvn->replace_node(cmpx, igvn->makecon(TypeInt::CC_EQ));
 864   igvn->rehash_node_delayed(call);
 865   call->del_req(call->req()-1);
 866 }
 867 
 868 void ShenandoahBarrierSetC2::enqueue_useful_gc_barrier(PhaseIterGVN* igvn, Node* node) const {
 869   if (node->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(node)) {
 870     igvn->add_users_to_worklist(node);
 871   }
 872 }
 873 
 874 void ShenandoahBarrierSetC2::eliminate_useless_gc_barriers(Unique_Node_List &useful, Compile* C) const {
 875   for (uint i = 0; i < useful.size(); i++) {
 876     Node* n = useful.at(i);
 877     if (n->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(n)) {
 878       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
 879         C->record_for_igvn(n->fast_out(i));
 880       }
 881     }
 882   }
 883   for (int i = state()->enqueue_barriers_count() - 1; i >= 0; i--) {
 884     ShenandoahEnqueueBarrierNode* n = state()->enqueue_barrier(i);
 885     if (!useful.member(n)) {
 886       state()->remove_enqueue_barrier(n);
 887     }
 888   }
 889   for (int i = state()->load_reference_barriers_count() - 1; i >= 0; i--) {
 890     ShenandoahLoadReferenceBarrierNode* n = state()->load_reference_barrier(i);
 891     if (!useful.member(n)) {
 892       state()->remove_load_reference_barrier(n);
 893     }
 894   }
 895 }
 896 
 897 void* ShenandoahBarrierSetC2::create_barrier_state(Arena* comp_arena) const {
 898   return new(comp_arena) ShenandoahBarrierSetC2State(comp_arena);
 899 }
 900 
 901 ShenandoahBarrierSetC2State* ShenandoahBarrierSetC2::state() const {
 902   return reinterpret_cast<ShenandoahBarrierSetC2State*>(Compile::current()->barrier_set_state());
 903 }
 904 
 905 // If the BarrierSetC2 state has kept macro nodes in its compilation unit state to be
 906 // expanded later, then now is the time to do so.
 907 bool ShenandoahBarrierSetC2::expand_macro_nodes(PhaseMacroExpand* macro) const { return false; }
 908 
 909 #ifdef ASSERT
 910 void ShenandoahBarrierSetC2::verify_gc_barriers(Compile* compile, CompilePhase phase) const {
 911   if (ShenandoahVerifyOptoBarriers && phase == BarrierSetC2::BeforeMacroExpand) {
 912     ShenandoahBarrierC2Support::verify(Compile::current()->root());
 913   } else if (phase == BarrierSetC2::BeforeCodeGen) {
 914     // Verify G1 pre-barriers
 915     const int marking_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_active_offset());
 916 
 917     ResourceArea *area = Thread::current()->resource_area();
 918     Unique_Node_List visited(area);
 919     Node_List worklist(area);
 920     // We're going to walk control flow backwards starting from the Root
 921     worklist.push(compile->root());
 922     while (worklist.size() > 0) {
 923       Node *x = worklist.pop();
 924       if (x == NULL || x == compile->top()) continue;
 925       if (visited.member(x)) {
 926         continue;
 927       } else {
 928         visited.push(x);
 929       }
 930 
 931       if (x->is_Region()) {
 932         for (uint i = 1; i < x->req(); i++) {
 933           worklist.push(x->in(i));
 934         }
 935       } else {
 936         worklist.push(x->in(0));
 937         // We are looking for the pattern:
 938         //                            /->ThreadLocal
 939         // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
 940         //              \->ConI(0)
 941         // We want to verify that the If and the LoadB have the same control
 942         // See GraphKit::g1_write_barrier_pre()
 943         if (x->is_If()) {
 944           IfNode *iff = x->as_If();
 945           if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
 946             CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
 947             if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
 948                 && cmp->in(1)->is_Load()) {
 949               LoadNode *load = cmp->in(1)->as_Load();
 950               if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
 951                   && load->in(2)->in(3)->is_Con()
 952                   && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
 953 
 954                 Node *if_ctrl = iff->in(0);
 955                 Node *load_ctrl = load->in(0);
 956 
 957                 if (if_ctrl != load_ctrl) {
 958                   // Skip possible CProj->NeverBranch in infinite loops
 959                   if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
 960                       && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
 961                     if_ctrl = if_ctrl->in(0)->in(0);
 962                   }
 963                 }
 964                 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
 965               }
 966             }
 967           }
 968         }
 969       }
 970     }
 971   }
 972 }
 973 #endif
 974 
 975 Node* ShenandoahBarrierSetC2::ideal_node(PhaseGVN* phase, Node* n, bool can_reshape) const {
 976   if (is_shenandoah_wb_pre_call(n)) {
 977     uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type()->domain()->cnt();
 978     if (n->req() > cnt) {
 979       Node* addp = n->in(cnt);
 980       if (has_only_shenandoah_wb_pre_uses(addp)) {
 981         n->del_req(cnt);
 982         if (can_reshape) {
 983           phase->is_IterGVN()->_worklist.push(addp);
 984         }
 985         return n;
 986       }
 987     }
 988   }
 989   if (n->Opcode() == Op_CmpP) {
 990     Node* in1 = n->in(1);
 991     Node* in2 = n->in(2);
 992     if (in1->bottom_type() == TypePtr::NULL_PTR) {
 993       in2 = step_over_gc_barrier(in2);
 994     }
 995     if (in2->bottom_type() == TypePtr::NULL_PTR) {
 996       in1 = step_over_gc_barrier(in1);
 997     }
 998     PhaseIterGVN* igvn = phase->is_IterGVN();
 999     if (in1 != n->in(1)) {
1000       if (igvn != NULL) {
1001         n->set_req_X(1, in1, igvn);
1002       } else {
1003         n->set_req(1, in1);
1004       }
1005       assert(in2 == n->in(2), "only one change");
1006       return n;
1007     }
1008     if (in2 != n->in(2)) {
1009       if (igvn != NULL) {
1010         n->set_req_X(2, in2, igvn);
1011       } else {
1012         n->set_req(2, in2);
1013       }
1014       return n;
1015     }
1016   } else if (can_reshape &&
1017              n->Opcode() == Op_If &&
1018              ShenandoahBarrierC2Support::is_heap_stable_test(n) &&
1019              n->in(0) != NULL) {
1020     Node* dom = n->in(0);
1021     Node* prev_dom = n;
1022     int op = n->Opcode();
1023     int dist = 16;
1024     // Search up the dominator tree for another heap stable test
1025     while (dom->Opcode() != op    ||  // Not same opcode?
1026            !ShenandoahBarrierC2Support::is_heap_stable_test(dom) ||  // Not same input 1?
1027            prev_dom->in(0) != dom) {  // One path of test does not dominate?
1028       if (dist < 0) return NULL;
1029 
1030       dist--;
1031       prev_dom = dom;
1032       dom = IfNode::up_one_dom(dom);
1033       if (!dom) return NULL;
1034     }
1035 
1036     // Check that we did not follow a loop back to ourselves
1037     if (n == dom) {
1038       return NULL;
1039     }
1040 
1041     return n->as_If()->dominated_by(prev_dom, phase->is_IterGVN());
1042   }
1043 
1044   return NULL;
1045 }
1046 
1047 bool ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(Node* n) {
1048   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1049     Node* u = n->fast_out(i);
1050     if (!is_shenandoah_wb_pre_call(u)) {
1051       return false;
1052     }
1053   }
1054   return n->outcnt() > 0;
1055 }
1056 
1057 bool ShenandoahBarrierSetC2::final_graph_reshaping(Compile* compile, Node* n, uint opcode) const {
1058   switch (opcode) {
1059     case Op_CallLeaf:
1060     case Op_CallLeafNoFP: {
1061       assert (n->is_Call(), "");
1062       CallNode *call = n->as_Call();
1063       if (ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(call)) {
1064         uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type()->domain()->cnt();
1065         if (call->req() > cnt) {
1066           assert(call->req() == cnt + 1, "only one extra input");
1067           Node *addp = call->in(cnt);
1068           assert(!ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(addp), "useless address computation?");
1069           call->del_req(cnt);
1070         }
1071       }
1072       return false;
1073     }
1074     case Op_ShenandoahCompareAndSwapP:
1075     case Op_ShenandoahCompareAndSwapN:
1076     case Op_ShenandoahWeakCompareAndSwapN:
1077     case Op_ShenandoahWeakCompareAndSwapP:
1078     case Op_ShenandoahCompareAndExchangeP:
1079     case Op_ShenandoahCompareAndExchangeN:
1080 #ifdef ASSERT
1081       if( VerifyOptoOopOffsets ) {
1082         MemNode* mem  = n->as_Mem();
1083         // Check to see if address types have grounded out somehow.
1084         const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
1085         ciInstanceKlass *k = tp->klass()->as_instance_klass();
1086         bool oop_offset_is_sane = k->contains_field_offset(tp->offset());
1087         assert( !tp || oop_offset_is_sane, "" );
1088       }
1089 #endif
1090       return true;
1091     case Op_ShenandoahLoadReferenceBarrier:
1092       assert(false, "should have been expanded already");
1093       return true;
1094     default:
1095       return false;
1096   }
1097 }
1098 
1099 bool ShenandoahBarrierSetC2::escape_add_to_con_graph(ConnectionGraph* conn_graph, PhaseGVN* gvn, Unique_Node_List* delayed_worklist, Node* n, uint opcode) const {
1100   switch (opcode) {
1101     case Op_ShenandoahCompareAndExchangeP:
1102     case Op_ShenandoahCompareAndExchangeN:
1103       conn_graph->add_objload_to_connection_graph(n, delayed_worklist);
1104       // fallthrough
1105     case Op_ShenandoahWeakCompareAndSwapP:
1106     case Op_ShenandoahWeakCompareAndSwapN:
1107     case Op_ShenandoahCompareAndSwapP:
1108     case Op_ShenandoahCompareAndSwapN:
1109       conn_graph->add_to_congraph_unsafe_access(n, opcode, delayed_worklist);
1110       return true;
1111     case Op_StoreP: {
1112       Node* adr = n->in(MemNode::Address);
1113       const Type* adr_type = gvn->type(adr);
1114       // Pointer stores in G1 barriers looks like unsafe access.
1115       // Ignore such stores to be able scalar replace non-escaping
1116       // allocations.
1117       if (adr_type->isa_rawptr() && adr->is_AddP()) {
1118         Node* base = conn_graph->get_addp_base(adr);
1119         if (base->Opcode() == Op_LoadP &&
1120           base->in(MemNode::Address)->is_AddP()) {
1121           adr = base->in(MemNode::Address);
1122           Node* tls = conn_graph->get_addp_base(adr);
1123           if (tls->Opcode() == Op_ThreadLocal) {
1124              int offs = (int) gvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
1125              const int buf_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
1126              if (offs == buf_offset) {
1127                return true; // Pre barrier previous oop value store.
1128              }
1129           }
1130         }
1131       }
1132       return false;
1133     }
1134     case Op_ShenandoahEnqueueBarrier:
1135       conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), delayed_worklist);
1136       break;
1137     case Op_ShenandoahLoadReferenceBarrier:
1138       conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahLoadReferenceBarrierNode::ValueIn), delayed_worklist);
1139       return true;
1140     default:
1141       // Nothing
1142       break;
1143   }
1144   return false;
1145 }
1146 
1147 bool ShenandoahBarrierSetC2::escape_add_final_edges(ConnectionGraph* conn_graph, PhaseGVN* gvn, Node* n, uint opcode) const {
1148   switch (opcode) {
1149     case Op_ShenandoahCompareAndExchangeP:
1150     case Op_ShenandoahCompareAndExchangeN: {
1151       Node *adr = n->in(MemNode::Address);
1152       conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL);
1153       // fallthrough
1154     }
1155     case Op_ShenandoahCompareAndSwapP:
1156     case Op_ShenandoahCompareAndSwapN:
1157     case Op_ShenandoahWeakCompareAndSwapP:
1158     case Op_ShenandoahWeakCompareAndSwapN:
1159       return conn_graph->add_final_edges_unsafe_access(n, opcode);
1160     case Op_ShenandoahEnqueueBarrier:
1161       conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), NULL);
1162       return true;
1163     case Op_ShenandoahLoadReferenceBarrier:
1164       conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahLoadReferenceBarrierNode::ValueIn), NULL);
1165       return true;
1166     default:
1167       // Nothing
1168       break;
1169   }
1170   return false;
1171 }
1172 
1173 bool ShenandoahBarrierSetC2::escape_has_out_with_unsafe_object(Node* n) const {
1174   return n->has_out_with(Op_ShenandoahCompareAndExchangeP) || n->has_out_with(Op_ShenandoahCompareAndExchangeN) ||
1175          n->has_out_with(Op_ShenandoahCompareAndSwapP, Op_ShenandoahCompareAndSwapN, Op_ShenandoahWeakCompareAndSwapP, Op_ShenandoahWeakCompareAndSwapN);
1176 
1177 }
1178 
1179 bool ShenandoahBarrierSetC2::matcher_find_shared_post_visit(Matcher* matcher, Node* n, uint opcode) const {
1180   switch (opcode) {
1181     case Op_ShenandoahCompareAndExchangeP:
1182     case Op_ShenandoahCompareAndExchangeN:
1183     case Op_ShenandoahWeakCompareAndSwapP:
1184     case Op_ShenandoahWeakCompareAndSwapN:
1185     case Op_ShenandoahCompareAndSwapP:
1186     case Op_ShenandoahCompareAndSwapN: {   // Convert trinary to binary-tree
1187       Node* newval = n->in(MemNode::ValueIn);
1188       Node* oldval = n->in(LoadStoreConditionalNode::ExpectedIn);
1189       Node* pair = new BinaryNode(oldval, newval);
1190       n->set_req(MemNode::ValueIn,pair);
1191       n->del_req(LoadStoreConditionalNode::ExpectedIn);
1192       return true;
1193     }
1194     default:
1195       break;
1196   }
1197   return false;
1198 }
1199 
1200 bool ShenandoahBarrierSetC2::matcher_is_store_load_barrier(Node* x, uint xop) const {
1201   return xop == Op_ShenandoahCompareAndExchangeP ||
1202          xop == Op_ShenandoahCompareAndExchangeN ||
1203          xop == Op_ShenandoahWeakCompareAndSwapP ||
1204          xop == Op_ShenandoahWeakCompareAndSwapN ||
1205          xop == Op_ShenandoahCompareAndSwapN ||
1206          xop == Op_ShenandoahCompareAndSwapP;
1207 }