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/shenandoahHeap.hpp"
  27 #include "gc/shenandoah/shenandoahHeuristics.hpp"
  28 #include "gc/shenandoah/shenandoahRuntime.hpp"
  29 #include "gc/shenandoah/shenandoahThreadLocalData.hpp"
  30 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
  31 #include "gc/shenandoah/c2/shenandoahSupport.hpp"
  32 #include "opto/arraycopynode.hpp"
  33 #include "opto/escape.hpp"
  34 #include "opto/graphKit.hpp"
  35 #include "opto/idealKit.hpp"
  36 #include "opto/macro.hpp"
  37 #include "opto/movenode.hpp"
  38 #include "opto/narrowptrnode.hpp"
  39 #include "opto/rootnode.hpp"
  40 
  41 ShenandoahBarrierSetC2* ShenandoahBarrierSetC2::bsc2() {
  42   return reinterpret_cast<ShenandoahBarrierSetC2*>(BarrierSet::barrier_set()->barrier_set_c2());
  43 }
  44 
  45 ShenandoahBarrierSetC2State::ShenandoahBarrierSetC2State(Arena* comp_arena)
  46   : _enqueue_barriers(new (comp_arena) GrowableArray<ShenandoahEnqueueBarrierNode*>(comp_arena, 8,  0, NULL)),
  47     _load_reference_barriers(new (comp_arena) GrowableArray<ShenandoahLoadReferenceBarrierNode*>(comp_arena, 8,  0, NULL)) {
  48 }
  49 
  50 int ShenandoahBarrierSetC2State::enqueue_barriers_count() const {
  51   return _enqueue_barriers->length();
  52 }
  53 
  54 ShenandoahEnqueueBarrierNode* ShenandoahBarrierSetC2State::enqueue_barrier(int idx) const {
  55   return _enqueue_barriers->at(idx);
  56 }
  57 
  58 void ShenandoahBarrierSetC2State::add_enqueue_barrier(ShenandoahEnqueueBarrierNode * n) {
  59   assert(!_enqueue_barriers->contains(n), "duplicate entry in barrier list");
  60   _enqueue_barriers->append(n);
  61 }
  62 
  63 void ShenandoahBarrierSetC2State::remove_enqueue_barrier(ShenandoahEnqueueBarrierNode * n) {
  64   if (_enqueue_barriers->contains(n)) {
  65     _enqueue_barriers->remove(n);
  66   }
  67 }
  68 
  69 int ShenandoahBarrierSetC2State::load_reference_barriers_count() const {
  70   return _load_reference_barriers->length();
  71 }
  72 
  73 ShenandoahLoadReferenceBarrierNode* ShenandoahBarrierSetC2State::load_reference_barrier(int idx) const {
  74   return _load_reference_barriers->at(idx);
  75 }
  76 
  77 void ShenandoahBarrierSetC2State::add_load_reference_barrier(ShenandoahLoadReferenceBarrierNode * n) {
  78   assert(!_load_reference_barriers->contains(n), "duplicate entry in barrier list");
  79   _load_reference_barriers->append(n);
  80 }
  81 
  82 void ShenandoahBarrierSetC2State::remove_load_reference_barrier(ShenandoahLoadReferenceBarrierNode * n) {
  83   if (_load_reference_barriers->contains(n)) {
  84     _load_reference_barriers->remove(n);
  85   }
  86 }
  87 
  88 Node* ShenandoahBarrierSetC2::shenandoah_storeval_barrier(GraphKit* kit, Node* obj) const {
  89   if (ShenandoahStoreValEnqueueBarrier) {
  90     obj = shenandoah_enqueue_barrier(kit, obj);
  91   }
  92   return obj;
  93 }
  94 
  95 #define __ kit->
  96 
  97 bool ShenandoahBarrierSetC2::satb_can_remove_pre_barrier(GraphKit* kit, PhaseTransform* phase, Node* adr,
  98                                                          BasicType bt, uint adr_idx) const {
  99   intptr_t offset = 0;
 100   Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
 101   AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
 102 
 103   if (offset == Type::OffsetBot) {
 104     return false; // cannot unalias unless there are precise offsets
 105   }
 106 
 107   if (alloc == NULL) {
 108     return false; // No allocation found
 109   }
 110 
 111   intptr_t size_in_bytes = type2aelembytes(bt);
 112 
 113   Node* mem = __ memory(adr_idx); // start searching here...
 114 
 115   for (int cnt = 0; cnt < 50; cnt++) {
 116 
 117     if (mem->is_Store()) {
 118 
 119       Node* st_adr = mem->in(MemNode::Address);
 120       intptr_t st_offset = 0;
 121       Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
 122 
 123       if (st_base == NULL) {
 124         break; // inscrutable pointer
 125       }
 126 
 127       // Break we have found a store with same base and offset as ours so break
 128       if (st_base == base && st_offset == offset) {
 129         break;
 130       }
 131 
 132       if (st_offset != offset && st_offset != Type::OffsetBot) {
 133         const int MAX_STORE = BytesPerLong;
 134         if (st_offset >= offset + size_in_bytes ||
 135             st_offset <= offset - MAX_STORE ||
 136             st_offset <= offset - mem->as_Store()->memory_size()) {
 137           // Success:  The offsets are provably independent.
 138           // (You may ask, why not just test st_offset != offset and be done?
 139           // The answer is that stores of different sizes can co-exist
 140           // in the same sequence of RawMem effects.  We sometimes initialize
 141           // a whole 'tile' of array elements with a single jint or jlong.)
 142           mem = mem->in(MemNode::Memory);
 143           continue; // advance through independent store memory
 144         }
 145       }
 146 
 147       if (st_base != base
 148           && MemNode::detect_ptr_independence(base, alloc, st_base,
 149                                               AllocateNode::Ideal_allocation(st_base, phase),
 150                                               phase)) {
 151         // Success:  The bases are provably independent.
 152         mem = mem->in(MemNode::Memory);
 153         continue; // advance through independent store memory
 154       }
 155     } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
 156 
 157       InitializeNode* st_init = mem->in(0)->as_Initialize();
 158       AllocateNode* st_alloc = st_init->allocation();
 159 
 160       // Make sure that we are looking at the same allocation site.
 161       // The alloc variable is guaranteed to not be null here from earlier check.
 162       if (alloc == st_alloc) {
 163         // Check that the initialization is storing NULL so that no previous store
 164         // has been moved up and directly write a reference
 165         Node* captured_store = st_init->find_captured_store(offset,
 166                                                             type2aelembytes(T_OBJECT),
 167                                                             phase);
 168         if (captured_store == NULL || captured_store == st_init->zero_memory()) {
 169           return true;
 170         }
 171       }
 172     }
 173 
 174     // Unless there is an explicit 'continue', we must bail out here,
 175     // because 'mem' is an inscrutable memory state (e.g., a call).
 176     break;
 177   }
 178 
 179   return false;
 180 }
 181 
 182 #undef __
 183 #define __ ideal.
 184 
 185 void ShenandoahBarrierSetC2::satb_write_barrier_pre(GraphKit* kit,
 186                                                     bool do_load,
 187                                                     Node* obj,
 188                                                     Node* adr,
 189                                                     uint alias_idx,
 190                                                     Node* val,
 191                                                     const TypeOopPtr* val_type,
 192                                                     Node* pre_val,
 193                                                     BasicType bt) const {
 194   // Some sanity checks
 195   // Note: val is unused in this routine.
 196 
 197   if (do_load) {
 198     // We need to generate the load of the previous value
 199     assert(obj != NULL, "must have a base");
 200     assert(adr != NULL, "where are loading from?");
 201     assert(pre_val == NULL, "loaded already?");
 202     assert(val_type != NULL, "need a type");
 203 
 204     if (ReduceInitialCardMarks
 205         && satb_can_remove_pre_barrier(kit, &kit->gvn(), adr, bt, alias_idx)) {
 206       return;
 207     }
 208 
 209   } else {
 210     // In this case both val_type and alias_idx are unused.
 211     assert(pre_val != NULL, "must be loaded already");
 212     // Nothing to be done if pre_val is null.
 213     if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
 214     assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
 215   }
 216   assert(bt == T_OBJECT, "or we shouldn't be here");
 217 
 218   IdealKit ideal(kit, true);
 219 
 220   Node* tls = __ thread(); // ThreadLocalStorage
 221 
 222   Node* no_base = __ top();
 223   Node* zero  = __ ConI(0);
 224   Node* zeroX = __ ConX(0);
 225 
 226   float likely  = PROB_LIKELY(0.999);
 227   float unlikely  = PROB_UNLIKELY(0.999);
 228 
 229   // Offsets into the thread
 230   const int index_offset   = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset());
 231   const int buffer_offset  = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
 232 
 233   // Now the actual pointers into the thread
 234   Node* buffer_adr  = __ AddP(no_base, tls, __ ConX(buffer_offset));
 235   Node* index_adr   = __ AddP(no_base, tls, __ ConX(index_offset));
 236 
 237   // Now some of the values
 238   Node* marking;
 239   Node* gc_state = __ AddP(no_base, tls, __ ConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset())));
 240   Node* ld = __ load(__ ctrl(), gc_state, TypeInt::BYTE, T_BYTE, Compile::AliasIdxRaw);
 241   marking = __ AndI(ld, __ ConI(ShenandoahHeap::MARKING));
 242   assert(ShenandoahBarrierC2Support::is_gc_state_load(ld), "Should match the shape");
 243 
 244   // if (!marking)
 245   __ if_then(marking, BoolTest::ne, zero, unlikely); {
 246     BasicType index_bt = TypeX_X->basic_type();
 247     assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size.");
 248     Node* index   = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
 249 
 250     if (do_load) {
 251       // load original value
 252       // alias_idx correct??
 253       pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
 254     }
 255 
 256     // if (pre_val != NULL)
 257     __ if_then(pre_val, BoolTest::ne, kit->null()); {
 258       Node* buffer  = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
 259 
 260       // is the queue for this thread full?
 261       __ if_then(index, BoolTest::ne, zeroX, likely); {
 262 
 263         // decrement the index
 264         Node* next_index = kit->gvn().transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
 265 
 266         // Now get the buffer location we will log the previous value into and store it
 267         Node *log_addr = __ AddP(no_base, buffer, next_index);
 268         __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
 269         // update the index
 270         __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
 271 
 272       } __ else_(); {
 273 
 274         // logging buffer is full, call the runtime
 275         const TypeFunc *tf = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type();
 276         __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), "shenandoah_wb_pre", pre_val, tls);
 277       } __ end_if();  // (!index)
 278     } __ end_if();  // (pre_val != NULL)
 279   } __ end_if();  // (!marking)
 280 
 281   // Final sync IdealKit and GraphKit.
 282   kit->final_sync(ideal);
 283 
 284   if (ShenandoahSATBBarrier && adr != NULL) {
 285     Node* c = kit->control();
 286     Node* call = c->in(1)->in(1)->in(1)->in(0);
 287     assert(is_shenandoah_wb_pre_call(call), "shenandoah_wb_pre call expected");
 288     call->add_req(adr);
 289   }
 290 }
 291 
 292 bool ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(Node* call) {
 293   return call->is_CallLeaf() &&
 294          call->as_CallLeaf()->entry_point() == CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry);
 295 }
 296 
 297 bool ShenandoahBarrierSetC2::is_shenandoah_lrb_call(Node* call) {
 298   return call->is_CallLeaf() &&
 299          call->as_CallLeaf()->entry_point() == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_JRT);
 300 }
 301 
 302 bool ShenandoahBarrierSetC2::is_shenandoah_marking_if(PhaseTransform *phase, Node* n) {
 303   if (n->Opcode() != Op_If) {
 304     return false;
 305   }
 306 
 307   Node* bol = n->in(1);
 308   assert(bol->is_Bool(), "");
 309   Node* cmpx = bol->in(1);
 310   if (bol->as_Bool()->_test._test == BoolTest::ne &&
 311       cmpx->is_Cmp() && cmpx->in(2) == phase->intcon(0) &&
 312       is_shenandoah_state_load(cmpx->in(1)->in(1)) &&
 313       cmpx->in(1)->in(2)->is_Con() &&
 314       cmpx->in(1)->in(2) == phase->intcon(ShenandoahHeap::MARKING)) {
 315     return true;
 316   }
 317 
 318   return false;
 319 }
 320 
 321 bool ShenandoahBarrierSetC2::is_shenandoah_state_load(Node* n) {
 322   if (!n->is_Load()) return false;
 323   const int state_offset = in_bytes(ShenandoahThreadLocalData::gc_state_offset());
 324   return n->in(2)->is_AddP() && n->in(2)->in(2)->Opcode() == Op_ThreadLocal
 325          && n->in(2)->in(3)->is_Con()
 326          && n->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == state_offset;
 327 }
 328 
 329 void ShenandoahBarrierSetC2::shenandoah_write_barrier_pre(GraphKit* kit,
 330                                                           bool do_load,
 331                                                           Node* obj,
 332                                                           Node* adr,
 333                                                           uint alias_idx,
 334                                                           Node* val,
 335                                                           const TypeOopPtr* val_type,
 336                                                           Node* pre_val,
 337                                                           BasicType bt) const {
 338   if (ShenandoahSATBBarrier) {
 339     IdealKit ideal(kit);
 340     kit->sync_kit(ideal);
 341 
 342     satb_write_barrier_pre(kit, do_load, obj, adr, alias_idx, val, val_type, pre_val, bt);
 343 
 344     ideal.sync_kit(kit);
 345     kit->final_sync(ideal);
 346   }
 347 }
 348 
 349 Node* ShenandoahBarrierSetC2::shenandoah_enqueue_barrier(GraphKit* kit, Node* pre_val) const {
 350   return kit->gvn().transform(new ShenandoahEnqueueBarrierNode(pre_val));
 351 }
 352 
 353 // Helper that guards and inserts a pre-barrier.
 354 void ShenandoahBarrierSetC2::insert_pre_barrier(GraphKit* kit, Node* base_oop, Node* offset,
 355                                                 Node* pre_val, bool need_mem_bar) const {
 356   // We could be accessing the referent field of a reference object. If so, when G1
 357   // is enabled, we need to log the value in the referent field in an SATB buffer.
 358   // This routine performs some compile time filters and generates suitable
 359   // runtime filters that guard the pre-barrier code.
 360   // Also add memory barrier for non volatile load from the referent field
 361   // to prevent commoning of loads across safepoint.
 362 
 363   // Some compile time checks.
 364 
 365   // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
 366   const TypeX* otype = offset->find_intptr_t_type();
 367   if (otype != NULL && otype->is_con() &&
 368       otype->get_con() != java_lang_ref_Reference::referent_offset) {
 369     // Constant offset but not the reference_offset so just return
 370     return;
 371   }
 372 
 373   // We only need to generate the runtime guards for instances.
 374   const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
 375   if (btype != NULL) {
 376     if (btype->isa_aryptr()) {
 377       // Array type so nothing to do
 378       return;
 379     }
 380 
 381     const TypeInstPtr* itype = btype->isa_instptr();
 382     if (itype != NULL) {
 383       // Can the klass of base_oop be statically determined to be
 384       // _not_ a sub-class of Reference and _not_ Object?
 385       ciKlass* klass = itype->klass();
 386       if ( klass->is_loaded() &&
 387           !klass->is_subtype_of(kit->env()->Reference_klass()) &&
 388           !kit->env()->Object_klass()->is_subtype_of(klass)) {
 389         return;
 390       }
 391     }
 392   }
 393 
 394   // The compile time filters did not reject base_oop/offset so
 395   // we need to generate the following runtime filters
 396   //
 397   // if (offset == java_lang_ref_Reference::_reference_offset) {
 398   //   if (instance_of(base, java.lang.ref.Reference)) {
 399   //     pre_barrier(_, pre_val, ...);
 400   //   }
 401   // }
 402 
 403   float likely   = PROB_LIKELY(  0.999);
 404   float unlikely = PROB_UNLIKELY(0.999);
 405 
 406   IdealKit ideal(kit);
 407 
 408   Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
 409 
 410   __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
 411       // Update graphKit memory and control from IdealKit.
 412       kit->sync_kit(ideal);
 413 
 414       Node* ref_klass_con = kit->makecon(TypeKlassPtr::make(kit->env()->Reference_klass()));
 415       Node* is_instof = kit->gen_instanceof(base_oop, ref_klass_con);
 416 
 417       // Update IdealKit memory and control from graphKit.
 418       __ sync_kit(kit);
 419 
 420       Node* one = __ ConI(1);
 421       // is_instof == 0 if base_oop == NULL
 422       __ if_then(is_instof, BoolTest::eq, one, unlikely); {
 423 
 424         // Update graphKit from IdeakKit.
 425         kit->sync_kit(ideal);
 426 
 427         // Use the pre-barrier to record the value in the referent field
 428         satb_write_barrier_pre(kit, false /* do_load */,
 429                                NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
 430                                pre_val /* pre_val */,
 431                                T_OBJECT);
 432         if (need_mem_bar) {
 433           // Add memory barrier to prevent commoning reads from this field
 434           // across safepoint since GC can change its value.
 435           kit->insert_mem_bar(Op_MemBarCPUOrder);
 436         }
 437         // Update IdealKit from graphKit.
 438         __ sync_kit(kit);
 439 
 440       } __ end_if(); // _ref_type != ref_none
 441   } __ end_if(); // offset == referent_offset
 442 
 443   // Final sync IdealKit and GraphKit.
 444   kit->final_sync(ideal);
 445 }
 446 
 447 #undef __
 448 
 449 const TypeFunc* ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type() {
 450   const Type **fields = TypeTuple::fields(2);
 451   fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value
 452   fields[TypeFunc::Parms+1] = TypeRawPtr::NOTNULL; // thread
 453   const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields);
 454 
 455   // create result type (range)
 456   fields = TypeTuple::fields(0);
 457   const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields);
 458 
 459   return TypeFunc::make(domain, range);
 460 }
 461 
 462 const TypeFunc* ShenandoahBarrierSetC2::shenandoah_clone_barrier_Type() {
 463   const Type **fields = TypeTuple::fields(1);
 464   fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value
 465   const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+1, fields);
 466 
 467   // create result type (range)
 468   fields = TypeTuple::fields(0);
 469   const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields);
 470 
 471   return TypeFunc::make(domain, range);
 472 }
 473 
 474 const TypeFunc* ShenandoahBarrierSetC2::shenandoah_load_reference_barrier_Type() {
 475   const Type **fields = TypeTuple::fields(1);
 476   fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value
 477   const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+1, fields);
 478 
 479   // create result type (range)
 480   fields = TypeTuple::fields(1);
 481   fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL;
 482   const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+1, fields);
 483 
 484   return TypeFunc::make(domain, range);
 485 }
 486 
 487 Node* ShenandoahBarrierSetC2::store_at_resolved(C2Access& access, C2AccessValue& val) const {
 488   DecoratorSet decorators = access.decorators();
 489 
 490   const TypePtr* adr_type = access.addr().type();
 491   Node* adr = access.addr().node();
 492 
 493   bool anonymous = (decorators & ON_UNKNOWN_OOP_REF) != 0;
 494   bool on_heap = (decorators & IN_HEAP) != 0;
 495 
 496   if (!access.is_oop() || (!on_heap && !anonymous)) {
 497     return BarrierSetC2::store_at_resolved(access, val);
 498   }
 499 
 500   GraphKit* kit = access.kit();
 501 
 502   uint adr_idx = kit->C->get_alias_index(adr_type);
 503   assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
 504   Node* value = val.node();
 505   value = shenandoah_storeval_barrier(kit, value);
 506   val.set_node(value);
 507   shenandoah_write_barrier_pre(kit, true /* do_load */, /*kit->control(),*/ access.base(), adr, adr_idx, val.node(),
 508                                static_cast<const TypeOopPtr*>(val.type()), NULL /* pre_val */, access.type());
 509   return BarrierSetC2::store_at_resolved(access, val);
 510 }
 511 
 512 Node* ShenandoahBarrierSetC2::load_at_resolved(C2Access& access, const Type* val_type) const {
 513   DecoratorSet decorators = access.decorators();
 514 
 515   Node* adr = access.addr().node();
 516   Node* obj = access.base();
 517 
 518   bool mismatched = (decorators & C2_MISMATCHED) != 0;
 519   bool unknown = (decorators & ON_UNKNOWN_OOP_REF) != 0;
 520   bool on_heap = (decorators & IN_HEAP) != 0;
 521   bool on_weak = (decorators & ON_WEAK_OOP_REF) != 0;
 522   bool is_unordered = (decorators & MO_UNORDERED) != 0;
 523   bool need_cpu_mem_bar = !is_unordered || mismatched || !on_heap;
 524 
 525   Node* top = Compile::current()->top();
 526 
 527   Node* offset = adr->is_AddP() ? adr->in(AddPNode::Offset) : top;
 528   Node* load = BarrierSetC2::load_at_resolved(access, val_type);
 529 
 530   if (access.is_oop()) {
 531     if (ShenandoahLoadRefBarrier) {
 532       load = new ShenandoahLoadReferenceBarrierNode(NULL, load);
 533       load = access.kit()->gvn().transform(load);
 534     }
 535   }
 536 
 537   // If we are reading the value of the referent field of a Reference
 538   // object (either by using Unsafe directly or through reflection)
 539   // then, if SATB is enabled, we need to record the referent in an
 540   // SATB log buffer using the pre-barrier mechanism.
 541   // Also we need to add memory barrier to prevent commoning reads
 542   // from this field across safepoint since GC can change its value.
 543   bool need_read_barrier = ShenandoahKeepAliveBarrier &&
 544     (on_heap && (on_weak || (unknown && offset != top && obj != top)));
 545 
 546   if (!access.is_oop() || !need_read_barrier) {
 547     return load;
 548   }
 549 
 550   GraphKit* kit = access.kit();
 551 
 552   if (on_weak) {
 553     // Use the pre-barrier to record the value in the referent field
 554     satb_write_barrier_pre(kit, false /* do_load */,
 555                            NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
 556                            load /* pre_val */, T_OBJECT);
 557     // Add memory barrier to prevent commoning reads from this field
 558     // across safepoint since GC can change its value.
 559     kit->insert_mem_bar(Op_MemBarCPUOrder);
 560   } else if (unknown) {
 561     // We do not require a mem bar inside pre_barrier if need_mem_bar
 562     // is set: the barriers would be emitted by us.
 563     insert_pre_barrier(kit, obj, offset, load, !need_cpu_mem_bar);
 564   }
 565 
 566   return load;
 567 }
 568 
 569 static void pin_atomic_op(C2AtomicAccess& access) {
 570   if (!access.needs_pinning()) {
 571     return;
 572   }
 573   // SCMemProjNodes represent the memory state of a LoadStore. Their
 574   // main role is to prevent LoadStore nodes from being optimized away
 575   // when their results aren't used.
 576   GraphKit* kit = access.kit();
 577   Node* load_store = access.raw_access();
 578   assert(load_store != NULL, "must pin atomic op");
 579   Node* proj = kit->gvn().transform(new SCMemProjNode(load_store));
 580   kit->set_memory(proj, access.alias_idx());
 581 }
 582 
 583 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_val_at_resolved(C2AtomicAccess& access, Node* expected_val,
 584                                                    Node* new_val, const Type* value_type) const {
 585   GraphKit* kit = access.kit();
 586   if (access.is_oop()) {
 587     new_val = shenandoah_storeval_barrier(kit, new_val);
 588     shenandoah_write_barrier_pre(kit, false /* do_load */,
 589                                  NULL, NULL, max_juint, NULL, NULL,
 590                                  expected_val /* pre_val */, T_OBJECT);
 591 
 592     MemNode::MemOrd mo = access.mem_node_mo();
 593     Node* mem = access.memory();
 594     Node* adr = access.addr().node();
 595     const TypePtr* adr_type = access.addr().type();
 596     Node* load_store = NULL;
 597 
 598 #ifdef _LP64
 599     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
 600       Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop()));
 601       Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop()));
 602       if (ShenandoahCASBarrier) {
 603         load_store = kit->gvn().transform(new ShenandoahCompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
 604       } else {
 605         load_store = kit->gvn().transform(new CompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
 606       }
 607     } else
 608 #endif
 609     {
 610       if (ShenandoahCASBarrier) {
 611         load_store = kit->gvn().transform(new ShenandoahCompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo));
 612       } else {
 613         load_store = kit->gvn().transform(new CompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo));
 614       }
 615     }
 616 
 617     access.set_raw_access(load_store);
 618     pin_atomic_op(access);
 619 
 620 #ifdef _LP64
 621     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
 622       load_store = kit->gvn().transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
 623     }
 624 #endif
 625     load_store = kit->gvn().transform(new ShenandoahLoadReferenceBarrierNode(NULL, load_store));
 626     return load_store;
 627   }
 628   return BarrierSetC2::atomic_cmpxchg_val_at_resolved(access, expected_val, new_val, value_type);
 629 }
 630 
 631 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_bool_at_resolved(C2AtomicAccess& access, Node* expected_val,
 632                                                               Node* new_val, const Type* value_type) const {
 633   GraphKit* kit = access.kit();
 634   if (access.is_oop()) {
 635     new_val = shenandoah_storeval_barrier(kit, new_val);
 636     shenandoah_write_barrier_pre(kit, false /* do_load */,
 637                                  NULL, NULL, max_juint, NULL, NULL,
 638                                  expected_val /* pre_val */, T_OBJECT);
 639     DecoratorSet decorators = access.decorators();
 640     MemNode::MemOrd mo = access.mem_node_mo();
 641     Node* mem = access.memory();
 642     bool is_weak_cas = (decorators & C2_WEAK_CMPXCHG) != 0;
 643     Node* load_store = NULL;
 644     Node* adr = access.addr().node();
 645 #ifdef _LP64
 646     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
 647       Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop()));
 648       Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop()));
 649       if (ShenandoahCASBarrier) {
 650         if (is_weak_cas) {
 651           load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
 652         } else {
 653           load_store = kit->gvn().transform(new ShenandoahCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
 654         }
 655       } else {
 656         if (is_weak_cas) {
 657           load_store = kit->gvn().transform(new WeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
 658         } else {
 659           load_store = kit->gvn().transform(new CompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
 660         }
 661       }
 662     } else
 663 #endif
 664     {
 665       if (ShenandoahCASBarrier) {
 666         if (is_weak_cas) {
 667           load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
 668         } else {
 669           load_store = kit->gvn().transform(new ShenandoahCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
 670         }
 671       } else {
 672         if (is_weak_cas) {
 673           load_store = kit->gvn().transform(new WeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
 674         } else {
 675           load_store = kit->gvn().transform(new CompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
 676         }
 677       }
 678     }
 679     access.set_raw_access(load_store);
 680     pin_atomic_op(access);
 681     return load_store;
 682   }
 683   return BarrierSetC2::atomic_cmpxchg_bool_at_resolved(access, expected_val, new_val, value_type);
 684 }
 685 
 686 Node* ShenandoahBarrierSetC2::atomic_xchg_at_resolved(C2AtomicAccess& access, Node* val, const Type* value_type) const {
 687   GraphKit* kit = access.kit();
 688   if (access.is_oop()) {
 689     val = shenandoah_storeval_barrier(kit, val);
 690   }
 691   Node* result = BarrierSetC2::atomic_xchg_at_resolved(access, val, value_type);
 692   if (access.is_oop()) {
 693     result = kit->gvn().transform(new ShenandoahLoadReferenceBarrierNode(NULL, result));
 694     shenandoah_write_barrier_pre(kit, false /* do_load */,
 695                                  NULL, NULL, max_juint, NULL, NULL,
 696                                  result /* pre_val */, T_OBJECT);
 697   }
 698   return result;
 699 }
 700 
 701 void ShenandoahBarrierSetC2::clone(GraphKit* kit, Node* src, Node* dst, Node* size, bool is_array) const {
 702   // TODO: Implement using proper barriers.
 703   BarrierSetC2::clone(kit, src, dst, size, is_array);
 704 }
 705 
 706 // Support for GC barriers emitted during parsing
 707 bool ShenandoahBarrierSetC2::is_gc_barrier_node(Node* node) const {
 708   if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) return true;
 709   if (node->Opcode() != Op_CallLeaf && node->Opcode() != Op_CallLeafNoFP) {
 710     return false;
 711   }
 712   CallLeafNode *call = node->as_CallLeaf();
 713   if (call->_name == NULL) {
 714     return false;
 715   }
 716 
 717   return strcmp(call->_name, "shenandoah_clone_barrier") == 0 ||
 718          strcmp(call->_name, "shenandoah_cas_obj") == 0 ||
 719          strcmp(call->_name, "shenandoah_wb_pre") == 0;
 720 }
 721 
 722 Node* ShenandoahBarrierSetC2::step_over_gc_barrier(Node* c) const {
 723   if (c->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
 724     return c->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
 725   }
 726   if (c->Opcode() == Op_ShenandoahEnqueueBarrier) {
 727     c = c->in(1);
 728   }
 729   return c;
 730 }
 731 
 732 bool ShenandoahBarrierSetC2::expand_barriers(Compile* C, PhaseIterGVN& igvn) const {
 733   return !ShenandoahBarrierC2Support::expand(C, igvn);
 734 }
 735 
 736 bool ShenandoahBarrierSetC2::optimize_loops(PhaseIdealLoop* phase, LoopOptsMode mode, VectorSet& visited, Node_Stack& nstack, Node_List& worklist) const {
 737   if (mode == LoopOptsShenandoahExpand) {
 738     assert(UseShenandoahGC, "only for shenandoah");
 739     ShenandoahBarrierC2Support::pin_and_expand(phase);
 740     return true;
 741   } else if (mode == LoopOptsShenandoahPostExpand) {
 742     assert(UseShenandoahGC, "only for shenandoah");
 743     visited.Clear();
 744     ShenandoahBarrierC2Support::optimize_after_expansion(visited, nstack, worklist, phase);
 745     return true;
 746   }
 747   return false;
 748 }
 749 
 750 bool ShenandoahBarrierSetC2::array_copy_requires_gc_barriers(BasicType type) const {
 751   return false;
 752 }
 753 
 754 bool ShenandoahBarrierSetC2::clone_needs_postbarrier(ArrayCopyNode *ac, PhaseIterGVN& igvn) {
 755   Node* src = ac->in(ArrayCopyNode::Src);
 756   const TypeOopPtr* src_type = igvn.type(src)->is_oopptr();
 757   if (src_type->isa_instptr() != NULL) {
 758     ciInstanceKlass* ik = src_type->klass()->as_instance_klass();
 759     if ((src_type->klass_is_exact() || (!ik->is_interface() && !ik->has_subklass())) && !ik->has_injected_fields()) {
 760       if (ik->has_object_fields()) {
 761         return true;
 762       } else {
 763         if (!src_type->klass_is_exact()) {
 764           igvn.C->dependencies()->assert_leaf_type(ik);
 765         }
 766       }
 767     } else {
 768       return true;
 769         }
 770   } else if (src_type->isa_aryptr()) {
 771     BasicType src_elem  = src_type->klass()->as_array_klass()->element_type()->basic_type();
 772     if (src_elem == T_OBJECT || src_elem == T_ARRAY) {
 773       return true;
 774     }
 775   } else {
 776     return true;
 777   }
 778   return false;
 779 }
 780 
 781 // Support for macro expanded GC barriers
 782 void ShenandoahBarrierSetC2::register_potential_barrier_node(Node* node) const {
 783   if (node->Opcode() == Op_ShenandoahEnqueueBarrier) {
 784     state()->add_enqueue_barrier((ShenandoahEnqueueBarrierNode*) node);
 785   }
 786   if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
 787     state()->add_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node);
 788   }
 789 }
 790 
 791 void ShenandoahBarrierSetC2::unregister_potential_barrier_node(Node* node) const {
 792   if (node->Opcode() == Op_ShenandoahEnqueueBarrier) {
 793     state()->remove_enqueue_barrier((ShenandoahEnqueueBarrierNode*) node);
 794   }
 795   if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
 796     state()->remove_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node);
 797   }
 798 }
 799 
 800 void ShenandoahBarrierSetC2::eliminate_gc_barrier(PhaseMacroExpand* macro, Node* n) const {
 801   if (is_shenandoah_wb_pre_call(n)) {
 802     shenandoah_eliminate_wb_pre(n, &macro->igvn());
 803   }
 804 }
 805 
 806 void ShenandoahBarrierSetC2::shenandoah_eliminate_wb_pre(Node* call, PhaseIterGVN* igvn) const {
 807   assert(UseShenandoahGC && is_shenandoah_wb_pre_call(call), "");
 808   Node* c = call->as_Call()->proj_out(TypeFunc::Control);
 809   c = c->unique_ctrl_out();
 810   assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
 811   c = c->unique_ctrl_out();
 812   assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
 813   Node* iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
 814   assert(iff->is_If(), "expect test");
 815   if (!is_shenandoah_marking_if(igvn, iff)) {
 816     c = c->unique_ctrl_out();
 817     assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
 818     iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
 819     assert(is_shenandoah_marking_if(igvn, iff), "expect marking test");
 820   }
 821   Node* cmpx = iff->in(1)->in(1);
 822   igvn->replace_node(cmpx, igvn->makecon(TypeInt::CC_EQ));
 823   igvn->rehash_node_delayed(call);
 824   call->del_req(call->req()-1);
 825 }
 826 
 827 void ShenandoahBarrierSetC2::enqueue_useful_gc_barrier(Unique_Node_List &worklist, Node* node) const {
 828   if (node->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(node)) {
 829     worklist.push(node);
 830   }
 831 }
 832 
 833 void ShenandoahBarrierSetC2::eliminate_useless_gc_barriers(Unique_Node_List &useful) const {
 834   for (uint i = 0; i < useful.size(); i++) {
 835     Node* n = useful.at(i);
 836     if (n->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(n)) {
 837       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
 838         Compile::current()->record_for_igvn(n->fast_out(i));
 839       }
 840     }
 841   }
 842   for (int i = state()->enqueue_barriers_count() - 1; i >= 0; i--) {
 843     ShenandoahEnqueueBarrierNode* n = state()->enqueue_barrier(i);
 844     if (!useful.member(n)) {
 845       state()->remove_enqueue_barrier(n);
 846     }
 847   }
 848   for (int i = state()->load_reference_barriers_count() - 1; i >= 0; i--) {
 849     ShenandoahLoadReferenceBarrierNode* n = state()->load_reference_barrier(i);
 850     if (!useful.member(n)) {
 851       state()->remove_load_reference_barrier(n);
 852     }
 853   }
 854 }
 855 
 856 void ShenandoahBarrierSetC2::add_users_to_worklist(Unique_Node_List* worklist) const {}
 857 
 858 void* ShenandoahBarrierSetC2::create_barrier_state(Arena* comp_arena) const {
 859   return new(comp_arena) ShenandoahBarrierSetC2State(comp_arena);
 860 }
 861 
 862 ShenandoahBarrierSetC2State* ShenandoahBarrierSetC2::state() const {
 863   return reinterpret_cast<ShenandoahBarrierSetC2State*>(Compile::current()->barrier_set_state());
 864 }
 865 
 866 // If the BarrierSetC2 state has kept macro nodes in its compilation unit state to be
 867 // expanded later, then now is the time to do so.
 868 bool ShenandoahBarrierSetC2::expand_macro_nodes(PhaseMacroExpand* macro) const { return false; }
 869 
 870 #ifdef ASSERT
 871 void ShenandoahBarrierSetC2::verify_gc_barriers(bool post_parse) const {
 872   if (ShenandoahVerifyOptoBarriers && !post_parse) {
 873     ShenandoahBarrierC2Support::verify(Compile::current()->root());
 874   }
 875 }
 876 #endif
 877 
 878 Node* ShenandoahBarrierSetC2::ideal_node(PhaseGVN* phase, Node* n, bool can_reshape) const {
 879   if (is_shenandoah_wb_pre_call(n)) {
 880     uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type()->domain()->cnt();
 881     if (n->req() > cnt) {
 882       Node* addp = n->in(cnt);
 883       if (has_only_shenandoah_wb_pre_uses(addp)) {
 884         n->del_req(cnt);
 885         if (can_reshape) {
 886           phase->is_IterGVN()->_worklist.push(addp);
 887         }
 888         return n;
 889       }
 890     }
 891   }
 892   return NULL;
 893 }
 894 
 895 bool ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(Node* n) {
 896   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
 897     Node* u = n->fast_out(i);
 898     if (!is_shenandoah_wb_pre_call(u)) {
 899       return false;
 900     }
 901   }
 902   return n->outcnt() > 0;
 903 }